Siesta: Recent developments and

applications
Cite as: J. Chem. Phys. 152, 204108 (2020); https://doi.org/10.1063/5.0005077
Submitted: 18 February 2020 . Accepted: 20 April 2020 . Published Online: 27 May 2020

Alberto García , Nick Papior , Arsalan Akhtar , Emilio Artacho , Volker Blum , Emanuele
Bosoni , Pedro Brandimarte, Mads Brandbyge, J. I. Cerdá, Fabiano Corsetti, Ramón Cuadrado,
Vladimir Dikan, Jaime Ferrer, Julian Gale, Pablo García-Fernández, V. M. García-Suárez , Sandra
García, Georg Huhs, Sergio Illera, Richard Korytár, Peter Koval , Irina Lebedeva, Lin Lin , Pablo
López-Tarifa , Sara G. Mayo, Stephan Mohr , Pablo Ordejón, Andrei Postnikov , Yann Pouillon,
Miguel Pruneda, Roberto Robles, Daniel Sánchez-Portal, Jose M. Soler , Rafi Ullah , Victor Wen-
zhe Yu, and Javier Junquera

COLLECTIONS

Paper published as part of the special topic on Collection and Electronic Structure Software
Note: This article is part of the JCP Special Topic on Electronic Structure Software.

This paper was selected as an Editor’s Pick

ARTICLES YOU MAY BE INTERESTED IN

The CRYSTAL code, 1976–2020 and beyond, a long story

The Journal of Chemical Physics 152, 204111 (2020); https://doi.org/10.1063/5.0004892

TURBOMOLE: Modular program suite for ab initio quantum-chemical and condensed-

matter simulations
The Journal of Chemical Physics 152, 184107 (2020); https://doi.org/10.1063/5.0004635

CP2K: An electronic structure and molecular dynamics software package - Quickstep:

Efficient and accurate electronic structure calculations
The Journal of Chemical Physics 152, 194103 (2020); https://doi.org/10.1063/5.0007045
J. Chem. Phys. 152, 204108 (2020); https://doi.org/10.1063/5.0005077 152, 204108

© 2020 Author(s).
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

SIESTA: Recent developments and applications

Cite as: J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077
Submitted: 18 February 2020 • Accepted: 20 April 2020 •
Published Online: 27 May 2020

Alberto García,1,a) Nick Papior,2,b) Arsalan Akhtar,3,c) Emilio Artacho,4,5,6,7,d) Volker Blum,8,9,e)
1,f) 5,g) 10,h) 11,i)
Emanuele Bosoni, Pedro Brandimarte, Mads Brandbyge, J. I. Cerdá, Fabiano Corsetti,4,j)
3,k) 1,l) 12,13,m) 14,n)
Ramón Cuadrado, Vladimir Dikan, Jaime Ferrer, Julian Gale, Pablo García-Fernández,15,o)
12,13,p) 3,q) 16,r)
V. M. García-Suárez, Sandra García, Georg Huhs, Sergio Illera,3,s) Richard Korytár,17,t)
18,u) 4,v) 19,20,w)
Peter Koval, Irina Lebedeva, Lin Lin, Pablo López-Tarifa,21,x) Sara G. Mayo,22,y)
16,z) 3,aa) 23,ab)
Stephan Mohr, Pablo Ordejón, Andrei Postnikov, Yann Pouillon,15,ac) Miguel Pruneda,3,ad)
21,ae) 5,21,af) 22,24,ag)
Roberto Robles, Daniel Sánchez-Portal, Jose M. Soler, Rafi Ullah,4,25,ah) Victor Wen-zhe Yu,8,ai)
15,aj)
and Javier Junquera

AFFILIATIONS
1
Institut de Ciència de Materials de Barcelona (ICMAB-CSIC), Bellaterra E-08193, Spain
2
DTU Computing Center, Technical University of Denmark, 2800 Kgs. Lyngby, Denmark
3
Catalan Institute of Nanoscience and Nanotechnology - ICN2, CSIC and BIST, Campus UAB, 08193 Bellaterra, Spain
4
CIC Nanogune BRTA, Tolosa Hiribidea 76, 20018 San Sebastián, Spain
5
Donostia International Physics Center (DIPC), Paseo Manuel de Lardizabal 4, 20018 Donostia-San Sebastian, Spain
6
Ikerbasque, Basque Foundation for Science, 48011 Bilbao, Spain
7
Theory of Condensed Matter, Cavendish Laboratory, University of Cambridge, Cambridge CB3 0HE, United Kingdom
8
Department of Mechanical Engineering and Materials Science, Duke University, Durham, North Carolina 27708, USA
9
Department of Chemistry, Duke University, Durham, North Carolina 27708, USA
10
DTU Physics, Center for Nanostructured Graphene (CNG), Technical University of Denmark, Kgs. Lyngby DK-2800, Denmark
11
Instituto de Ciencia de Materiales de Madrid ICMM-CSIC, Cantoblanco, 28049 Madrid, Spain
12
Department of Physics, University of Oviedo, Oviedo 33007, Spain
13
Nanomaterials and Nanotechnology Research Center, CSIC - Universidad de Oviedo, Oviedo 33007, Spain
14
Curtin Institute for Computation, Institute for Geoscience Research (TIGeR), School of Molecular and Life Sciences,
Curtin University, P.O. Box U1987, Perth, WA 6845, Australia
15
Departamento de Ciencias de la Tierra y Física de la Materia Condensada, Universidad de Cantabria,
Cantabria Campus Internacional, Avenida de los Castros s/n, 39005 Santander, Spain
16
Barcelona Supercomputing Center, c/Jordi Girona, 29, 08034 Barcelona, Spain
17
Department of Condensed Matter Physics, Faculty of Mathematics and Physics, Charles University, Ke Karlovu 5,
121 16 Praha 2, Czech Republic
18
Simune Atomistics S.L., Tolosa Hiribidea, 76, 20018 Donostia-San Sebastian, Spain
19
Department of Mathematics, University of California, Berkeley, California 94720, USA
20
Computational Research Division, Lawrence Berkeley National Laboratory, Berkeley, California 94720, USA
21
Centro de Física de Materiales, Centro Mixto CSIC-UPV/EHU, Paseo Manuel de Lardizabal 5, 20018
Donostia-San Sebastian, Spain
22
Departamento de Física de la Materia Condensada, Universidad Autónoma de Madrid, 28049 Madrid, Spain
23
LCP-A2MC, Université de Lorraine, 1 Bd Arago, F-57078 Metz, France
24
Instituto de Física de la Materia Condensada (IFIMAC), Universidad Autónoma de Madrid, 28049 Madrid, Spain
25
Departamento de Física de Materiales, UPV/EHU, Paseo Manuel de Lardizabal 3, 20018 Donostia-San Sebastián, Spain

Note: This article is part of the JCP Special Topic on Electronic Structure Software.
a)
Author to whom correspondence should be addressed: albertog@icmab.es
b)
Electronic mail: nicpa@dtu.dk

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-1

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

c)
Electronic mail: arsalan.akhtar@icn2.cat
d)
Electronic mail: ea245@cam.ac.uk
e)
Electronic mail: volker.blum@duke.edu
f)
Electronic mail: ebosoni@icmab.es
g)
Electronic mail: pedro_brandimarte001@ehu.eus
h)
Electronic mail: mabr@dtu.dk
i)
Electronic mail: jcerda@icmm.csic.es
j)
Electronic mail: fabiano.corsetti@gmail.com
k)
Electronic mail: ramon.cuadrado@gmail.com
l)
Electronic mail: vdikan@icmab.es
m)
Electronic mail: ferrer@uniovi.es
n)
Electronic mail: J.Gale@curtin.edu.au
o)
Electronic mail: garciapa@unican.es
p)
Electronic mail: vm.garciasuarez@gmail.com
q)
Electronic mail: sandragil@gmail.com
r)
Electronic mail: ghuhs@physik.hu-berlin.de
s)
Electronic mail: sergiollera22@gmail.com
t)
Electronic mail: korytar@karlov.mff.cuni.cz
u)
Electronic mail: koval.peter@gmail.com
v)
Electronic mail: i.lebedeva@nanogune.eu
w)
Electronic mail: linlin@math.berkeley.edu
x)
Electronic mail: pablolopeztarifa@gmail.com
y)
Electronic mail: sara.garciamayo@uam.es
z)
Electronic mail: stephan.mohr@bsc.es
aa)
Electronic mail: pablo.ordejon@icn2.cat
ab)
Electronic mail: andrei.postnikov@univ-lorraine.fr
ac)
Electronic mail: yann.pouillon@unican.es
ad)
Electronic mail: miguel.pruneda@icn2.cat
ae)
Electronic mail: roberto.robles@ehu.eus
af)
Electronic mail: daniel.sanchez@ehu.eus
ag)
Electronic mail: jose.soler@uam.es
ah)
Electronic mail: ullah1@llnl.gov
ai)
Electronic mail: wenzhe.yu@duke.edu
aj)
Electronic mail: javier.junquera@unican.es

ABSTRACT
A review of the present status, recent enhancements, and applicability of the SIESTA program is presented. Since its debut in the mid-1990s,
SIESTA’s flexibility, efficiency, and free distribution have given advanced materials simulation capabilities to many groups worldwide. The core
methodological scheme of SIESTA combines finite-support pseudo-atomic orbitals as basis sets, norm-conserving pseudopotentials, and a real-
space grid for the representation of charge density and potentials and the computation of their associated matrix elements. Here, we describe
the more recent implementations on top of that core scheme, which include full spin–orbit interaction, non-repeated and multiple-contact
ballistic electron transport, density functional theory (DFT)+U and hybrid functionals, time-dependent DFT, novel reduced-scaling solvers,
density-functional perturbation theory, efficient van der Waals non-local density functionals, and enhanced molecular-dynamics options. In
addition, a substantial effort has been made in enhancing interoperability and interfacing with other codes and utilities, such as WANNIER90 and
the second-principles modeling it can be used for, an AiiDA plugin for workflow automatization, interface to Lua for steering SIESTA runs, and
various post-processing utilities. SIESTA has also been engaged in the Electronic Structure Library effort from its inception, which has allowed
the sharing of various low-level libraries, as well as data standards and support for them, particularly the PSeudopotential Markup Language
definition and library for transferable pseudopotentials, and the interface to the ELectronic Structure Infrastructure library of solvers. Code
sharing is made easier by the new open-source licensing model of the program. This review also presents examples of application of the
capabilities of the code, as well as a view of on-going and future developments.
Published under license by AIP Publishing. https://doi.org/10.1063/5.0005077., s

I. INTRODUCTION in many disciplines. The SIESTA method and its implementation have
been key in this development, offering an efficient and flexible simu-
The possibility of treating large systems with first-principles lation paradigm based on the use of strictly localized basis sets. This
electronic structure methods has opened up new research avenues approach enables the implementation of reduced scaling algorithms,

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-2

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

and its accuracy and cost can be tuned in a wide range from quick The orbitals in the SIESTA basis set are made of the product of
exploratory calculations to highly accurate simulations, matching a real spherical harmonic and a radial function, which is numer-
the quality of other approaches, such as plane-wave (PW) methods. ically tabulated in a grid. The shape of the radial part is, in prin-
The SIESTA method has been described in detail in Ref. 1 with ciple, totally arbitrary, but the experience accumulated has proven
an update in Ref. 2. In this paper, we shall describe its present sta- that the numerical solution of the Schrodinger equation for a con-
tus, highlighting its strengths and documenting the steps that have fined isolated atom with the corresponding pseudopotential is a
recently been taken to improve its capabilities, performance, ease of very good choice in terms of accuracy vs computational cost. Fuller
use, and visibility in the electronic structure community. descriptions of the mechanisms to generate and optimize these
As we shall see, the improvements touch many areas. We pseudo-atomic orbitals (PAOs) are given in Refs. 8–10.
can underline the implementation of new core electronic structure The auxiliary real-space grid is an essential ingredient of the
features [density functional theory (DFT)+U, spin–orbit interac- method as it allows the efficient representation of charge densities
tion, and hybrid functionals], modes of operation [improved time- and potentials as well as the computation of the matrix elements of
dependent density functional theory (TD-DFT), density functional the Hamiltonian that cannot be handled as two-center integrals. This
perturbation theory (DFPT)], and analysis methods and procedures grid can be seen as the reciprocal space of a set of plane waves, and its
to access new properties. A major effort has been spent in enhanc- fineness is most conveniently parameterized by an energy cutoff (the
ing the interoperability of the code at various levels [sharing of “density” cutoff of plane-wave methods). There are limits to the soft-
pseudopotentials (PPs), a new wannierization interface opening the ness of the functions that can be described with such a grid, so core
way to sophisticated post-processing, and an interface to multiscale electrons are not considered (although semi-core electrons usually
methods]. Very significant performance enhancements have been are), and their effect is incorporated into pseudopotentials. The real-
made, notably to the TRANSIESTA module through improved algo- space grid is also used to solve the Poisson equation involved in the
rithms and to the core electronic structure problem through the computation of the electrostatic potential from the charge density
development of interfaces to new solvers. These advances have put through the use of a fast-Fourier-transform method. This means that
SIESTA in a prominent place in the high-performance electronic struc- SIESTA uses periodic boundary conditions (PBC). Non-periodic sys-
ture simulation scene, a role reinforced by its participation in impor- tems, such as molecules, tubes, or slabs, are treated using appropriate
tant international initiatives and by its new open-source licensing supercells.
model. SIESTA is now a mature code with more than 20 years of exis-
This manuscript is organized as follows: We provide an tence. In this period, the most important algorithms behind our
overview of the underlying methodology and the capabilities of implementation have been already fully described and documented
SIESTA in Sec. II, which serves to place the code in the wider ecosystem in a series of papers. Readers interested in the details of how the
of electronic structure materials simulation. Section III presents the basic elements defining the method are combined, as well as other
recent developments in and around the code, which are covered in relevant implementation details that make the method practical, can
Subsections III A–III O. To demonstrate SIESTA’s utility in the context find them in Ref. 1 and in the update with the new capabilities of the
of electronic structure calculations, we briefly present some relevant code in Ref. 2.
applications and survey a few areas in which SIESTA is being profitably We note that the term SIESTA is regularly used to describe both
used in Sec. IV. Plans for the future evolution of SIESTA are outlined the method (as outlined in Refs. 5 and 6) and its implementation in
in Sec. V. a computer program. The SIESTA method is at the basis of later inde-
pendent implementations, such as OpenMX11 and QuantumATK.12
Other subsequent codes are built on the method by revising some
II. KEY CONCEPTS OF SIESTA of the fundamental ingredients. This is the case of FHI-aims,13
which uses a more sophisticated real-space grid (atom-centered),
A. Theory background and context
thus extending the core scheme to all-electron calculations.
SIESTA appeared as a consequence of the push for linear-scaling In this paper, we describe new additions to the SIESTA code
electronic structure methods of the mid-1990s, which has been based on independent methodological advances, either pre-existent
reviewed, for example, in Refs. 3 and 4. SIESTA was the first linear- or specifically developed for SIESTA, as specified and cited in the
scaling self-consistent implementation of density functional theory appropriate sections below.
(DFT).5,6
The SIESTA method relies on atomic-like functions of finite B. Overview of SIESTA capabilities
support as basis sets7,8 —of arbitrary number, angular momentum, As a general purpose implementation, SIESTA can provide the
radial shape, and centers—combined with a discretization of space standard functionality available in mainstream DFT codes (ener-
for the computation of the Kohn–Sham (KS) Hamiltonian terms gies, forces, molecular-dynamics simulations, band structures, and
that involve more than two centers. The electron–ion interaction is densities of states) and shares with those codes the basic current
represented by norm-conserving pseudopotentials. These key ingre- limitations of DFT (notably the description of strongly correlated
dients, through the optimized handling of sparse matrices, are used systems).
to compute the self-consistent Hamiltonian and overlap matrices What makes SIESTA different from most other codes is the
with a computational expense that scales linearly with system size. atomic-like and strictly localized character of its basis set, which is
The method is completed with a choice of solvers for that Hamilto- at the root of its key strengths. The use of a “good first approxima-
nian, from optimized (but cube-scaling) diagonalization methods to tion” to the full problem implies, first, that a much smaller number of
reduced-scaling solvers of different flavors. basis functions are needed. Second, the finite-support of the orbitals

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-3

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

leads to sparsity and the possibility to use reduced-scaling methods. efficiently in modest hardware while also being able to exploit
Thus, high performance emerges almost by default. massive levels of parallelism in large supercomputers (see Fig. 1).
First, consider the basis cardinality: the number of basis orbitals It is also worth noting that the atomic character of the basis
per atom in a typical SIESTA calculation is of the order of 10–20. This set enables the use of a very intuitive suite of analysis tools since
is to be compared with a few hundred in the typical plane-wave most of the concepts relating to chemical bonding use the language
(PW) calculation. Furthermore, for systems whose description needs of atomic orbitals. Hence, SIESTA has a natural advantage in this
a vacuum region (e.g., slabs for surface calculations and 2D mono- area. Partial densities of states and atomic and crystal populations
layers), empty space is essentially “free” for SIESTA, whereas PW codes (COOP/COHP) are routinely used to gain insights into the stability
still need a basis set determined by the total size of the simulation and other properties of materials. For a recent example, see Ref. 14.
cell. SIESTA is then quite capable of dealing with systems composed of Similarly, an atomic basis provides a very natural and adequate lan-
dozens to hundreds of atoms on modest hardware, even when using guage for the first-principles simulation of electronic ballistic trans-
cubic-scaling diagonalization solvers, which are the default as they port in nanosized systems via the Green’s function-based Keldysh
are universally applicable. formalism implemented in TRANSIESTA,15 a part of the SIESTA package.
Electronic structure solvers with a more favorable size-scaling A very high number of citations of the SIESTA papers testify
can be applied to suitable systems. For example, one of SIESTA’s earlier to the successful application of the code to widely different sys-
calculations, in 1996, was a linear-scaling run for a strand of DNA tems. With regard to specific capabilities and the levels of accuracy
with 650 atoms, performed on a desktop workstation of the era.6 achievable, we can distinguish several levels. First, SIESTA implements
Reduced size-scaling is also a feature of the PEXSI solver described DFT, one of the most versatile materials simulation frameworks.
in Sec. III G 1 and of the NTPoly solver mentioned in Sec. III G 2. In DFT has its shortcomings, notably in regard to the description
addition to time-to-solution efficiency, these solvers have a smaller of strongly correlated systems, but these are being addressed (see
memory footprint than diagonalization, as the relevant matrices are Secs. III C and III E). Second, SIESTA uses pseudopotentials to rep-
kept in a sparse form rather than converting them to a dense format. resent the electron–ion interaction. The pseudopotential approach
Crucially, SIESTA’s baseline efficiency can be scaled up to ever- is firmly rooted in a sound physical approximation (that bond-
larger systems by parallelization. Both distributed [message-passing ing effects depend mostly on the valence electrons); however, it
(MPI)] and shared-memory (OpenMP) parallelization options are is at a disadvantage when core-electron effects are important (but
implemented in the code. As will be shown by some examples see Sec. III N 6). Third, SIESTA employs periodic boundary con-
in Sec. IV, non-trivial calculations with thousands of atoms are ditions (PBC) for the solution of the Poisson problem, sharing
used in applications in different contexts from molecular biology to with plane-wave codes the need to resort to repeated supercells
electronic transport. for the study of low-dimensional systems and to special techniques
Work on the performance aspects of the code is continuous, for the treatment of charged systems. However, it is important to
mostly on the solvers, which usually consume most of the computer note that, unlike plane-wave codes, SIESTA is only bound to PBC
time due to the very high efficiency of the Hamiltonian setup mod- because of the present treatment of the Hartree term of the single-
ule in SIESTA. This task is facilitated (see Sec. III O) by leveraging particle Hamiltonian. This limitation is addressed by the incorpora-
external libraries and developments generated by a number of inter- tion of alternative Poisson solvers, as described in Sec. III O, which
national initiatives in which SIESTA participates. The code can still run allow for open boundary conditions, as for isolated nano-systems,

FIG. 1. Parallel strong scaling of

SIESTA-PEXSI and the (Scala-
pack) diagonalization approach for
a DNA chain and a graphene–boron
nitride stack, prototypes of large
(hundreds of thousands of orbitals)
quasi-one-dimensional (1D) and quasi-
two-dimensional (2D) systems. “ppp”
stands for the number of MPI processes
used in each pole computation, and “sp”
the sparsity of the Hamiltonian (for more
details, see Sec. III G 1).

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-4

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

and hybrid open/periodic boundary conditions in different dimen- III. RECENT DEVELOPMENTS IN SIESTA
sions, as for isolated wires and slabs. It should be remembered
A. New distribution model and development
that the three approximations mentioned in this paragraph are very
infrastructure
widely used in the community, shared by some of the most popular
electronic structure codes. A few years ago, in 2016, a decision was made to change the
Fourth, with regard to SIESTA-specific approximations, partic- licensing model for SIESTA: traditionally, it had always been free of
ularly the basis set, it should be stressed that SIESTA is limited to charge to academics, but non-academic use required a special license
basis sets composed of functions that are the product of a radial and redistribution was not permitted. Now, SIESTA is formally an
part and spherical harmonics, but it does not constrain on how open-source program distributed according to the terms of the GNU
many, where such functions are centered, and the size of their General Public License (GPL).18 At the same time, the development
finite-support region. Calculations can flexibly range from quick infrastructure was made more transparent and scalable using the
exploration to very high-quality simulations (one may recall that Launchpad platform19 earlier and the Gitlab service now.20 The net
accuracy gold standards in the electronic structure are provided by effect of the changes has been a more fertile and dynamic devel-
quantum-chemistry methods based on LCAO). opment, with more contributors who can have direct access to the
The use of an atomic-orbital basis set implies however the limi- various branches of development and a better experience for users
tation of non-uniformity of convergence. As opposed to plane-wave who can download codes and raise issues in an integrated platform.
methods in which a single energy cutoff parameter monotonically These changes have been substantial for the core developers,
determines the quality of the calculation, there is no univocal proce- and the transitory period is still being felt. The main code base
dure for the choice of an appropriate basis set. This is a well-known is gradually gaining new developments that were planned long in
problem shared by the whole quantum-chemistry community, and advance, and new ones made possible by the greater openness
there is widely used and tested know-how. As shown in Fig. 2, it is and fluidity of the development model. Most of the new features
possible to attain, in practice, accuracy comparable to that of well- described below are already part of public releases, but a few are
converged plane-wave calculations. The reader is also referred to undergoing the last stages of testing before release. The work-flow
Secs. IV B and IV C 1 for the demonstrated examples of the accuracy is also moving from long-lived releases, hard to maintain with bug-
of the code, among many others in the literature. fixes, to more frequent releases that will be maintained for a shorter
To close this section, we stress that it has been a traditional and time.
deliberate attitude by the SIESTA team that, although proposing sen-
sible starting points to users as defaults, the choice of fundamental
B. New pseudopotential format for interoperability
approximations and inputs to the program (not only basis sets but
also density functionals and pseudopotentials) is a responsibility of PSML (for PSeudopotential Markup Language)21,22 is a file for-
the users, who retain full control and the flexibility to adapt the code mat for norm-conserving pseudopotential data, which is designed
to their specific needs. Nevertheless, tools for basis optimization are to encapsulate as much as possible the abstract concepts involved
provided with the program, new curated databases of pseudopoten- in the domain and to provide appropriate metadata and provenance
tials are coming online, and new ways to ameliorate the correlation information. This extra level of formalization aims at removing the
problem are being implemented. Some of these developments are interoperability problems associated with bespoke pseudopotential
described in subsequent sections. formats, which usually were designed to serve the needs of specific
generators and client codes and thus contain implicit assumptions
about the meaning of the data or lack information not considered
relevant.
PSML files can be produced by the ONCVPSP23 and ATOM24 pseu-
dopotential generator programs and are a download-format option
in the Pseudo-Dojo database of curated pseudopotentials.25,26
The software library libPSML21,22 can be used by electronic
structure codes to transparently extract the information in a PSML
file and incorporate it into their own data structures or to cre-
ate converters for other formats. It is currently used by SIESTA and
17,27
ABINIT, making possible a full pseudopotential interoperability
and facilitating comparisons of calculation results.
The use of this new format opens the door to benefit from
the availability of a Periodic Table of reliable and accurate norm-
conserving pseudopotentials, easing in most cases the task of pseu-
dopotential quality control.

FIG. 2. Basis set convergence for the binding energy (E b ) of a water dimer. Details
can be found in Ref. 16. The horizontal dotted line represents the converged plane- C. DFT+U for correlated systems
wave (PW) calculation (1300 eV cutoff) for the same system (dimer geometry and
The LDA+U method, initially developed by Anisimov and
box), pseudopotentials and density functional, using the ABINIT code.17 Inset:
deviation of E b vs the PW reference. The deviation for the last point is of 10 μeV.
co-workers28 with the objective to improve the treatment of the
electron–electron interaction for localized electrons within the bare

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-5

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

LDA description, has been implemented in SIESTA. The idea behind 3.08 eV [from the bare Generalized-Gradient-Approximation–
the LDA+U consists in describing the “strongly correlated” elec- Perdew–Burke–Ernzerhof (GGA-PBE) value of 1.08 eV], very close
tronic states of a system (typically, localized d or f orbitals) using to the experimental value for the onset of optical absorption in NiO32
the Hubbard model, whereas the rest of valence electrons are treated (3.10 eV). The magnetic moment on the Ni atom is also prop-
at the level of “standard” approximate DFT functionals.29 In the erly described with a value of 1.67 μB , which lies well within the
current version of SIESTA, the implementation is based on the sim- experimental range of values (between 1.64 μB 33 and 1.9 μB 34 ), and
plified rotationally invariant functional proposed by Dudarev and improves on the result of 1.39 μB obtained with a bare GGA–PBE
co-workers.30 Here, the corrections are made invariant under rota- functional.
tion of the atomic orbitals used to define the occupation number of
the correlated subspace, at the cost of retaining only the lowest order D. van der Waals functionals
Slater integrals in the factorization of the integrals of the Coulomb
kernel of the electron–electron interaction and neglecting the higher An efficient calculation of van der Waals (vdW) functionals35,36
order ones (i.e., taking the exchange interaction J as 0). The expres- was developed and first implemented in SIESTA using a polynomial
sion of the corrective term as a function of the occupation number expansion in the local variables (q1 , q2 ) of the nonlocal interaction
nIσ kernel Φ(q1 , q2 , r12 ) and a Fourier expansion in the relative position
ℓm of the localized correlated orbital ℓm with spin σ within the atom
I is given by r12 .37 As a result, the scaling of the vdW computation decreases from
O(N 2 ) to O(N log N), and it becomes marginal within the overall
U Iℓ cost. This scheme was later extended16 to a more complex kernel38
EU = ∑ [∑ nIσ Iσ
ℓm (1 − nℓm )], (1)
Iσℓ 2 m of the form Φ(n1 , |∇n1 |, n2 , |∇n2 |, r12 ), and it has been applied to
a large variety of systems, such as carbon nanotubes,37 hydrogen
where only one interaction parameter U I ℓ is needed to specify the adsorption,39,40 or liquid water.41
interaction per atom and ℓ-shell. In the practical SIESTA implementa-
tion, the populations on the correlated orbitals are computed using
non-overlapping (i.e., orthogonal) localized projectors. They can E. Hybrid functionals
be generated using either (i) the same algorithm used to produce The screened hybrid functional HSE0642–44 has been imple-
the first-ζ orbitals of the basis set, but with a larger energy shift, mented in SIESTA building on the work of Ref. 45. This functional
or (ii) cutting the exact solution of the pseudoatom with a Fermi is the result of adding nonlocal Hartree–Fock type exact exchange
function. (HFX) into semilocal density functionals. The Coulomb potential
The results of the LDA+U method are sensitively dependent on that appears in the exchange interaction is screened, so it has a
the numerical value of the effective on-site electronic interaction, the shorter range than 1/r. Here, to reduce the big prefactor involved
Hubbard U. Although, in principle, the value of U can be computed in the computation of the HFX potential matrix elements, we fit the
from first principles using linear response methods,31 a common NAO of the basis set with Gaussian-type orbitals, especially suited
practice is to tune it semiempirically, seeking agreement of certain to computing the four center electron repulsion integrals (ERIs) in a
properties (for instance, bandgaps or magnetic moments) with the straightforward and efficient analytical way. An example of this fit-
available experimental measurements. Then, the fitted U is used in ting for the 2s and 2p atomic orbitals basis set of oxygen is shown
subsequent calculations to predict other properties. in Fig. 4. The LIBINT package46 is required to calculate primitive
The LDA+U corrects localized states for which the self- ERIs, where recursive schemes of the Obara–Saika47 method and
interaction correction is expected to be stronger and is an effective the Head-Gordon and Pople’s variation48 thereof are implemented.
method to improve the description of the (underestimated) bandgap ERIs are calculated in the first self-consistent-field (SCF) cycle and
of insulators, as shown in Fig. 3 for the case of NiO. Once the Hub- then stored on disk. Only the ERIs with non-negligible contributions
bard correction is switched on, the optical bandgap increases up to are calculated, keeping the HFX Hamiltonian also sparse.

FIG. 3. Band structure of NiO in the

undistorted rock-salt type structure with
rhombohedral symmetry introduced
by a type-II antiferromagnetic order.
The experimental lattice spacing is
used. The bands obtained within
the GGA–Perdew–Burke–Ernzerhof
functional [panel (a)] and with a Hubbard
U correction of 4.6 eV applied on the
d-orbitals of Ni [panel (b)], as in Ref. 31,
are shown. The zero of the energy is set
at the top of the valence band.

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-6

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

FIG. 4. Gaussian fits of the radial part of oxygen 2s (a) and 2p (b) orbitals using six Gaussian functions. The orbitals to fit are represented by blue dots and the corresponding
Gaussian expansions by green continuous lines. Dashed vertical lines represent the standard deviations of individual Gaussians, and a red continuous line marks their upper
limit. The orbitals are set to zero in the yellow area, marking their cutoff radii.

This HSE06 functional has been used to compute the band F. Spin–orbit coupling
structure of bulk Si [diamond structure, Fig. 5(a)] and BaTiO3 [cubic The capability to include the spin–orbit (SO) interaction in
structure, Fig. 5(b)] with a double-zeta polarized basis set at the SIESTA and in the analysis tools is seen as a strategic asset for the
equilibrium lattice constant of the Perdew–Burke–Ernzerhof func- project in view of the recent interest in topological insulators and
tional49 within the generalized gradient approximation (5.499 Å for quasi-two-dimensional systems with important spin–orbit effects,
Si and 4.033 Å for BaTiO3 ). In both cases, the gap is opened with such as some of the transition metal dichalcogenides. It also brings
respect to the value obtained with the semilocal functional. In bulk the possibility to obtain the magnetic crystalline anisotropy (MCA)
Si, the bandgap is indirect: the top of the valence band is located (change in the total energy of the system upon changing the spin
at Γ and the bottom of the conduction band at a point along the quantization axis).
Γ → X high-symmetry line. It increases from 0.64 eV within GGA In a standard collinear-spin DFT calculation, the total KS
to 1.00 eV with the hybrid functional, in good agreement with Hamiltonian is represented by two independent spin-blocks, Ĥμν σσ
the experimental value of 1.17 eV.50 For the case of the perovskite [σ = ↑, ↓]. However, when the SO coupling is included, off-diagonal
oxide BaTiO3 , the bandgap is also indirect, from R to Γ, and its spin blocks arise (i.e., there are non-zero couplings between the two
value increases from 1.87 eV with the GGA to 3.28 eV with the spin components). Therefore, similar to the non-collinear spin case,
HSE06 functional, almost agreeing the experimental value of 3.2 eV the Hamiltonian becomes a full 2 × 2 matrix in spin space,
estimated by Wemple in the cubic phase.51

FIG. 5. Band structure of (a) bulk Si

in the diamond structure and (b) bulk
BaTiO3 in the cubic structure obtained
with the Perdew–Burke–Ernzerhof func-
tional (red lines) and with the HSE06
hybrid functional (black lines). The zero
of energies has been set to the valence
band maximum.

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-7

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

↑↑ ↑↓
KS ⎛Ĥμν Ĥμν ⎞ problem HΦ = εSΦ). The stage devoted to the calculation of the
Ĥμν = , (2) Hamiltonian H and overlap S is typically much lighter weight, as
↓↑
⎝Ĥμν ↓↓ ⎠
Ĥμν those matrices are intrinsically sparse due to the use of a finite-
support basis set. Accordingly, SIESTA’s performance is almost com-
where μν subindexes refer to the SIESTA basis orbitals. The fully rel- pletely linked to the use of appropriate external solver libraries.
ativistic (FR) Hamiltonian Ĥ KS is expressed as a sum of the kinetic Over the past few years, we have expanded the choices avail-
energy T̂, the scalar-relativistic pseudopotential part in the form of able to users and refined the relevant interfaces. Initially, we added
Kleinman–Bylander projectors V̂ KB , the spin–orbit V̂ SO term, and support for new individual solvers as detailed below, but recently
the Hartree V̂ H and exchange–correlation V̂ XC potentials, we have consolidated some of the most important functionality
under a new common interface to the ELSI (ELectronic Structure
Ĥ KS = T̂ + V̂ KB + V̂ SO + V̂ H + V̂ XC . (3)
Infrastructure) library of solvers.57,58
The first three terms of the right-hand side do not depend on 1. Solvers with a native interface
the charge density, ρ(r), and therefore do not change in the self-
Diagonalization (solution of the generalized eigenproblem
consistent cycle, while V̂ SO and V̂ XC are the only spin-dependent
appropriate for non-orthogonal orbitals) is the default method for
terms that couple both spin components.
obtaining the density-matrix in SIESTA. A number of standard rou-
In order to compute the MCAs, different orientations of the
tines are contained in the SCALAPACK library,59 but more effi-
spin quantization axis need to be considered. This may be done by
cient alternatives are possible. In particular, the ELPA library60–62
rotating either V̂ SO (as done by Cuadrado and Cerdá52 ) or the den-
uses an extra intermediate step in the tridiagonal conversion of the
sity matrix, which is the approach currently followed by SIESTA for
matrices to obtain better scalability and significant speedups over
compatibility with the non-collinear case.
SCALAPACK. An interface to ELPA is offered in SIESTA, so this solver
In the current implementation, the SO term is included non-
can be used as a drop-in replacement for SCALAPACK throughout
perturbatively so that the fully relativistic Hamiltonian is solved
the code.
self-consistently after extending the Kohn–Sham wavefunctions to
In addition, SIESTA has implemented interfaces to several meth-
full spinors. The following two different approaches have been
ods not based on diagonalization. In most cases, the use of a finite-
implemented in SIESTA to account for the SO term, V̂ SO :
support basis set, leading to the appearance of sparse matrices, is a
● on-site approximation: significant factor to achieve good performance:
Based on the work of Fernández-Seivane et al.,53,54 only
● The Fermi Operator Expansion (FOE) method63 uses the
the intra-atomic SO contributions within each l-shell of
formal relationship between Hamiltonian and density-
each atom are considered. In this approach, the SO terms
matrix, ρ̂ = fFD (Ĥ − μ), where fFD is the Fermi–Dirac func-
are obtained from analytically simple expressions for the
tion. A simple polynomial expansion of fFD can then be used
angular integrals, while the radial integrals are computed
to obtain ρ̂ without diagonalization. This method is imple-
numerically.
mented in the CheSS library,64 developed within the BigDFT
● off-site approach:
project.65
Here, V̂ SO is built following the Hemstreet formalism52,55
● The PEXSI method66,67 uses a pole expansion of fFD to get ρ̂
whereby a fully relativistic pseudopotential (FR-PP) opera-
in the following form:
tor is constructed in a fully separable form, i.e., non-local in
the radial part as well as in the angular variables, in order to P ρ
ωl
substantially reduce the computational cost. The necessary ρ̂ = Im(∑ ), (4)
lj Kleinman–Bylander projectors may be either constructed l=1 H − (zl + μ)S
by SIESTA itself from relativistic semilocal PPs or directly read
ρ
from appropriately generated PSML files, as provided by the where ωl and zl are the weights and poles for the corre-
Pseudo-Dojo project.25,26 Moreover, we note that the FR- sponding expansion of the Fermi–Dirac function, respec-
PP formalism (as well as the original one implemented in tively. The number of poles needed is significantly smaller
Ref. 52) uses the correct normalization constants Cl ±1/2 , in than for the polynomial version of the FOE, as its depen-
contrast with what was erroneously stated in Ref. 56. dence on the spectrum size is only logarithmic.
It would appear that having to invert matrices would
Although we consider the off-site approach more accurate, as
still render this approach cubic-scaling, but, in fact, only
it includes inter-shell and inter-atomic SO couplings, both approxi-
selected elements of ρ̂ have to be actually computed. This
mations yield very similar results in most of the tested systems with
“pole expansion and selected inversion” method offers a
relevant qualitative differences only found in a few specific cases.
SO reduced complexity (at most O(N 2 ) for dense systems and
Furthermore, the construction of the Vμν matrix is very fast under
O(N) for quasi-one-dimensional systems) and trivial par-
both schemes and involves a tiny fraction of the entire self-consistent
allelization over poles, so it is well-suited for very large
calculation.
problems on large machines. For example,68 computed the
electronic structure of large (up to 11 700 atoms) graphene
G. New electronic structure solvers
nanoflakes using SIESTA-PEXSI.
For most problems, SIESTA spends the largest fraction of CPU- ● The electronic structure problem can also be cast as a
time in the solver stage (solution of the generalized eigenvalue minimization problem (of an extended functional) without

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-8

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

orthogonalization. When additional localization constraints

are put in place, the original linear-scaling method in SIESTA
results. Without the extra localization constraints, the cubic-
scaling Orbital Minimization Method (OMM)69 can be com-
petitive with respect to diagonalization, as data can be reused
across SCF-cycle steps.

2. The ELSI interface

We have considerably extended the range of solver choices
and the performance enhancement possibilities of the code with
the integration of the open-source ELSI library (https://elsi-
interchange.org), which provides a unified software interface that
connects electronic structure codes to various high-performance
solver libraries to solve or circumvent eigenproblems encountered
in electronic structure theory.57 ELSI also ships with its own tested
versions of the individual solver libraries, but additionally, linking
against already compiled upstream versions from each solver library
is supported as much as possible.
The ELPA, OMM, and PEXSI solvers, which had their own
ad hoc interfaces as described in Sec. III G 1, are now available
through ELSI, which also supports other conventional dense eigen-
solvers (EigenExa70 and MAGMA71 ), sparse iterative eigensolvers
(SLEPc72 ), and linear scaling density matrix purification methods
(NTPoly73 ). As sketched in Fig. 6, an electronic structure code inter-
facing to ELSI automatically has access to all the eigensolvers and
density matrix solvers supported in ELSI. In addition, the ELSI
interface is able to convert arbitrarily distributed dense and sparse
matrices to the specification expected by the solvers, taking this
burden away from the electronic structure code. A comprehensive FIG. 6. Interaction of the ELSI interface with electronic structure codes. ELSI
review of the capabilities in the latest version of ELSI, including serves as a bridge between electronic structure codes and solver libraries.
An electronic structure code has access to various eigensolvers and density
parallel solution of problems found in spin-polarized systems (two matrix solvers via the ELSI API. Whenever necessary, ELSI handles the con-
spin channels) and periodic systems (multiple k-points), scalable version between different units, conventions, matrix formats, and programming
matrix I/O, density matrix extrapolation, iterative eigensolvers in languages.
a reverse communication interface (RCI) framework, has recently
been completed.58
With the common interface in place, any additions and
enhancements to the supported solvers can be used in SIESTA with
almost no code changes. This is particularly relevant for perfor-
mance enhancements, for example,
● Further levels of parallelization: A feature common, in
principle, to all solvers is that the SIESTA–ELSI interface
can exploit the full parallelization over k-points and spins
mentioned above. This means that these calculations can
use two extra levels of parallelization in the solver step
beyond the standard one of parallelization over orbitals (see
Fig. 7).
● The new version of the PEXSI solver integrated in ELSI can
achieve the same level of precision with fewer poles and
offers an extra level of parallelization over trial points for the
determination of the chemical-potential.
● Mixed-precision support: The ELPA solver can be invoked FIG. 7. Performance improvement from the use of the extra level of parallelization
in single-precision mode, which can speed up the ini- over k-points in SIESTA using the ELSI interface with the ELPA solver, compared to
tial steps of the electronic self-consistent-field (SCF) the previous diagonalization scheme (using both the standard SCALAPACK solver
cycle. and the existing ELPA interface in SIESTA). The system is bulk Si with H impurities,
with 1040 atoms, 13 328 orbitals, and a sampling of 8 k-points. The multi-k scheme
● Accelerator offloading: The ELPA library offers Graphical-
is able to stay closer to ideal scalability for larger numbers of MPI processes.
Processing-Unit (GPU) support in some kernels,62 and there

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-9

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

is scope for extending it to more kernels. ELSI also offers A way of taking such evolution into account in the discretized
an interface to the accelerator-enabled MAGMA library. implementation was proposed by Tomfohr and Sankey76 and is
Finally, the PEXSI developers are working on adding GPU based on a Löwdin orthonormalization. The scheme consists of two
support to the solver. steps. First, the wavefunctions are propagated using both S and H at
time t 0 using Eq. (7), but to an auxiliary coefficient matrix c̃,
H. Time dependent DFT Δt −1 Δt
c̃(t0 + Δt) = [S + iH(t0 ) ] [S − iH(t0 ) ]c(t0 ). (10)
Time-dependent density-functional theory (TD-DFT) was first 2 2
implemented in SIESTA in its real-time propagating form. It was first
described in Ref. 74 and then briefly in Ref. 2. It was based on Then, the propagation is followed by a change of basis operation
the Crank–Nicolson algorithm by which the effect of the evolution (only needed if the ionic positions have changed),
operator for an infinitesimal time step
c(t0 + Δt) = S− 2 (t0 + Δt) S 2 (t0 )c̃(t0 + Δt).
1 1
(11)
Û(t0 + Δt, t0 ) = exp[−iĤ(t)Δt] (5)
This algorithm is unitary by construction, and so the preser-
vation of orthonormality is guaranteed, regardless of the size of Δt.
on the wavefunction coefficients matrix at a given time t 0 , c(t 0 ) is
As discussed in detail in Ref. 75, this algorithm can be shown not
approximated by
to be entirely consistent with the connection represented by the
D matrix defined above. Nevertheless, the discrepancies due to the
Δt −1 Δt mentioned inconsistency have been shown to be small in a series
c(t0 + Δt) = [S + iH(t0 + Δt) ] [S − iH(t0 ) ]c(t0 ), (6)
2 2 of studies using this formalism,77–80 at least for low atomic veloc-
ities. The practical benefit of separating the two procedures is to
where Δt represents the finite time step resulting from time dis- perform the change of basis only when necessary, allowing for many
cretization, and S and H represent the overlap and Hamilto- electronic steps per atomic motion step, if the nuclei are still signif-
nian matrices, respectively, in the representation given by a non- icantly slower than electrons, for instance. The implementation of
orthogonal basis set, as used by SIESTA. That expression is obtained the Crank–Nicolson part is the same for both the fixed and moving
from equating the first-order evolution of the coefficients forward, basis.
from t 0 to t 0 + Δt/2, to the backward evolution from t 0 + Δt to the The square root and inverse square root are calculated by first
same intermediate step. computing its eigenvalues and eigenvectors,
It can be further simplified to
S = U E U†, (12)
−1
Δt Δt
c(t0 + Δt) = [S + iH(t0 ) ] [S − iH(t0 ) ]c(t0 ) (7) where E is a diagonal matrix with the eigenvalues of S. U is a square
2 2
matrix with the eigenvectors of S as its columns. Then,
for a smooth-enough variation of the Hamiltonian itself and a small
enough Δt, thereby avoiding the self-consistency implied in prop- S1/2 = U E1/2 U † and S−1/2 = U E−1/2 U † ,
agation using Eq. (6). In Sec. III H 3, recent developments on effi-
cient treatments of Eq. (6) beyond Eq. (7) are presented. Here, where E1/2 and E−1/2 are obtained by replacing diagonal elements
we describe the parallelization and related features in the TD-DFT of E with their square root and inverse square root (in the latter
implementation now found in standard SIESTA releases. case, neglecting those eigenvalues below a certain threshold value),
The implemented propagation is based on Eq. (7) (with the respectively.
improvement possibilities described in Sec. III H 3), but proper con- The two-stage algorithm has been implemented in SIESTA in par-
sideration must be taken of the fact that not only the coefficients allel, allowing for k-point sampling and for collinear spin. The initial
change in time but also the basis set and the Hilbert space spanned by occupied states to be propagated are read from a file. SIESTA is pre-
it when the atoms move. An analysis of the geometrical implications pared to run a conventional DFT calculation of whatever relevant
of this fact is presented in Ref. 75. The time-dependent Kohn–Sham initial state and write a wavefunction continuation file that acts as
equation initialization of the ulterior SIESTA run in real-time TD-DFT mode.
As it stands, SIESTA evolves states defined as fully occupied; partial
H∣ψ⟩ = i∂t ∣ψ⟩ (8) occupations are not currently supported.

becomes 1. Parallelization
Hc = i S (∂t + D)c, (9) The two-step procedure described above requires matrix–
matrix and matrix–scalar multiplication, matrix addition, and
where H, S, and c are the Hamiltonian, overlap, and coefficients matrix inversion, plus the diagonalization of the overlap matrix for
matrices, respectively, as before, and the D matrix is the con- the Löwdin step. Since only the occupied states are propagated, the
nection in the manifold given by the evolving Hilbert space,75 c matrix is rectangular N × N, that is, the number of propagating
Dμν = ⟨ϕμ |∂ t ϕν ⟩, for ϕμ and ϕν basis functions. states × the number of basis functions, while S and H are square,

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-10

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

N × N. The computation of the overlap and Hamiltonian matri-

ces is handled by pre-existing SIESTA routines, which are already
parallelized and well-optimized for High-Performance-Computing
(HPC) environments.69,81
The parallelization of the propagation following Eqs. (10) and
(11) is done simply by exploiting the MatrixSwitch library,82 which
allows for an abstracted manipulation of matrices, the details of
parallelization, data formats, conversions, etc., being taken care
of underneath. In this case, MatrixSwitch is called to use the
BLACS83 and SCALAPACK84 libraries, meaning that this part of
the code is run on the dense-matrix infrastructure, as already
done with conventional diagonalization solvers. As for the lat-
ter, although the H and S matrices are sparse, the c matrix is
dense.
Conversion between storage formats is an important consider-
ation here. The native matrix storage format employed by SIESTA is
a compressed sparse column (CSC) scheme with a one-dimensional
block-cyclic distribution (1D-BCD) over MPI processes. A block-
cyclic distribution is needed by the BLACS and SCALAPACK pack-
FIG. 8. The relative share of the total running time for the Crank–Nicolson algo-
age. The matrices can therefore be temporarily converted from rithm, the Löwdin step, and the rest of the program operations (including the
sparse to dense using the same parallel distribution; this is a very building of the SCF Hamiltonian) for a system of 5000 Ge atoms and one He
efficient operation, since no MPI communication is necessary. It projectile using 30–316 processors.
should be noted that a two-dimensional (2D) BCD is known to be
more efficient in terms of parallel scaling.69 The conversion from 1D
to 2D does however carry a heavier cost, as MPI communication is
inevitable. more (1 a.u. ∼ c/137, with c being the speed of light). In that case,
The parallel efficiency of our implementation is therefore the diagonalization step may be avoided (or replaced by the N × N
chiefly determined by that of the underlying SCALAPACK drivers. diagonalization of the overlap matrix for the evolving states, instead
The matrix inversion in Eq. (10) is performed using LU factor- of the N × N for the basis set overlap).
ization. For the diagonalization of the overlap matrix, we have For fast moving atoms within Ehrenfest dynamics, there is also
implemented the option of using either a standard diagonalization the need to implement correction terms to the forces related to both
approach (tridiagonal reduction followed by the implicit QR algo- the change of basis and the rotation of the time-dependent Hilbert
rithm) or a divide-and-conquer algorithm, as described in Ref. 85. space (the intrinsic curvature of the manifold). These terms are well
The latter is known to scale better with system size. known,86 and their geometrical meaning in terms of the relevant
The scaling with the number of processors is very similar to the curvature75 will be the subject of a future publication. They are being
scaling of a conventional DFT SIESTA run using diagonalization as tested and should be incorporated shortly to a visible branch in
the solver option, since both procedures are run on routines of anal- the open source repository, to be later merged into the trunk, and
ogous scaling within the same dense-matrix-algebra library. Figure 8 further incorporated into SIESTA releases.
shows the relative share in the total running time of the three main For efficiency, iterative inversion options will be explored
procedures involved: the Crank–Nicolson algorithm, the change of replacing the present LU implementation in SCALAPACK, and quite
basis, and the calculation of the SCF Hamiltonian plus other minor a few possibilities exist to incorporate more sophisticated algorithms
processes in SIESTA such as building the density matrix. This was per- to the described operations. What has been described is robust and
formed for a system of 5000 Ge + 1 He atoms described with a quite transparent, but the MatrixSwitch abstraction should allow
single-zeta polarized basis set. The Crank–Nicolson algorithm takes easy implementation of other techniques.
about 18% of the total time on 30 processors, which increases to
25% on 316 processors. Instead, the change of basis procedure takes
about 38% of the total time on 30 processors, which decreases as the 3. Improved real-time propagators
parallelization increases, reflecting its better scaling properties. The
Löwdin step is the most expensive operation on all numbers of pro- Equation (7) represents a fast approach of electronic prop-
cessors, which affects TD-DFT simulations (and only those steps) agation in real-time TD-DFT, especially suited to study systems
involving atomic motion. where the perturbation of the electronic density is relatively small
(e.g., optical linear response74 ). If one is interested in simulat-
2. TD-DFT beyond the released version ing systems with heavily perturbed electronic densities by exter-
There are many possible (and feasible) improvements on what nal forces (such as those exerted by intense laser fields or fast
has been described above, some of them in the pipeline. From a fun- atom collisions for instance), one should choose a more elabo-
damental point of view, the Löwdin step will be replaced by another rate propagation scheme that preserves better the time-reversibility
basis-changing step in the direction of what was proposed in Ref. 75. of the propagator operator. Some of the authors introduced in
It is needed if atoms move at velocities of around 1 atomic unit or Ref. 79 an extrapolation algorithm to study the stopping power of

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-11

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

prototype semiconductors. Briefly, the method uses Eq. (7) with an or carbon beams.91 This particular simulation addresses the frag-
extrapolated Hamiltonian, mentation pattern of a doubly ionized uracil molecule (its deep-
est Kohn–Sham orbital is empty) using the Becke- Lee-Yang-Parr
Δt −1 Δt (BLYP) density functional.92,93 The technical details for the simula-
c(t0 + Δt) = [S + iHext ] [S − iHext ]c(t0 ), (13)
2 2 tion with the CPMD code used as a reference can be found in Refs. 89
and 90. SIESTA calculations using the same density functional and the
where the extrapolated Hamiltonian H ext reads integrators described above use a double-zeta-polarized (DZP) basis
set. As shown in Fig. 9, the standard SIESTA implementation cannot
1
Hext = H(t0 ) + ΔH (14) properly deal with such a highly excited system. The extrapolation
2 scheme in Eq. (13) already provides a large improvement and gives
and an energy conservation similar to the CPMD simulations in Refs. 89
and 90. Finally, the two-step algorithm further improves the energy
ΔH = H(t0 ) − H(t0 − Δt). (15) conservation. For smaller time steps, the improvements given by the
Additionally, the user is given the option to divide each propagation two-step scheme are even more clear, as shown in Fig. 10.
step Δt into n sub-steps in an effort to increase the accuracy of the
first-order expansion underlying the derivation of Eq. (13). In this 4. Electronic stopping of atomic projectiles
case, the final equation reads
Let us finish the TD-DFT section with a brief mention of its
n −1 successful application to the problem of simulating the excitation
j Δt j Δt
c(t0 + Δt) = ∏[S + iHext ] [S − iHext ]c(t0 ), (16) of the electrons of a condensed matter system when traversed by a
j=1 2n 2n
high-energy atomic projectile (so-called electronic stopping, since
the electrons slow down the projectile). This physical problem is very
with
relevant to questions of interest to the nuclear and aerospace indus-
j 1 1 tries, as well as to the treatment of cancer. Despite its great relevance
Hext = H(t0 ) + ( j − )ΔH. (17)
n 2 and having been researched since Rutherford’s experiment in 1911,
the understanding of electronic stopping processes has been essen-
Recently, we introduced a third algorithm for propagation.
tially limited to either weak effects in the linear-response regime or
Leaving aside in this description the complications associated with
beyond linear but only for target systems close to the homogeneous
the possible subdivision of each time step, the new algorithm is
electron liquid (jellium).
based on a two-step scheme where the electronic wavefunction is
An earlier version of the TD-DFT implementation in SIESTA74
first propagated until half of the step, Δt/2, using extrapolation as in
allowed the first explicit first-principles simulation of electronic
Eq. (13),
stopping for protons and antiprotons in LiF, a wide-bandgap insula-
Δt Δt −1 Δt tor, which was quite successful.94 The difference of sign between pro-
c(t0 + ) = [S + iHext ] [S − iHext ]c(t0 ), (18) tons and antiprotons produces a significant difference in the stop-
2 4 4
ping power (rate of energy excitation) beyond the linear-response
then an explicit calculation of the half-step Hamiltonian, H(t 0 paradigm (the Barkas effect), and the insulating character of the
+ Δt/2), is performed using the coefficients c(t 0 + Δt/2) obtained
from Eq. (18). In a second step, the coefficients are evolved from
the beginning of the step, c(t 0 ), to the full step, c(t 0 + Δt), using the
half-step Hamiltonian,

Δt Δt −1 Δt Δt
c(t0 + Δt) = [S + iH(t0 + ) ] [S − iH(t0 + ) ]c(t0 ).
2 2 2 2
(19)

This approach, although increasing the CPU time by around ∼35%,

as compared to the two previous schemes, allows for better energy
conservation for highly perturbed systems where the Kohn–Sham
potential heavily varies in time.
In order to provide a more quantitative comparison between
the three schemes, namely, the default propagation of Eq. (7), the
extrapolation propagation of Eq. (13), and the two-step propaga-
tion of Eqs. (18) and (19), we compare their performance vs the
P-TDDFT implementation87 of the CPMD code88 in the case of a
double ionization of a uracil molecule in the gas phase.89,90 Simula-
tions of this type provide access to the ultrafast electronic dynamics FIG. 9. Energy drift in an energy conserving TDDFT simulation of ionized-core
uracil for the three propagation methods described here, as compared with the
that occurs at the atto and femto timescales in the ionized genetic
CPMD implementation, for a time step of Δt = 0.24 as (0.01 a.u.).
material (DNA and RNA) as a consequence of collisions with proton

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-12

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

FIG. 10. Energy drift comparison as in

Fig. 9 for time steps Δt = 0.24 and 1.2
as.

target makes it inaccessible to the jellium paradigm. The success similar computational costs, which are comparable to those obtained
stimulated further studies along this line77–80 using improved ver- with a finite difference approach. Figure 12 shows a comparison
sions of TD-DFT in SIESTA, as described here. Figure 11 displays between both methods for model fullerene-type systems of different
the electron deformation density around a proton displacing in a sizes.
bulk Ge target.79 They were also followed by analogous simulations
using plane-wave codes by an increase in the number of groups (for
a review, see Ref. 95), although the latter calculations do demand
considerably larger computational resources.

I. Density functional perturbation theory

The original implementation of density functional perturba-
tion theory, as a post-processing and independent code (LINRES96 ),
has been recently merged into SIESTA. It allows us to compute the
phonon dispersions using a supercell approach (Γ-point phonons).
Both LDA and GGA functionals can be used (through LibXC). Cal-
culation of the perturbed Hamiltonian and overlap matrix elements
follows the same methodology as for ground-state calculations with

FIG. 12. (Left panel) Comparison of the performance of the new DFPT approach
with the conventional finite-differences method in SIESTA. The time required to com-
pute a whole row of the dynamical matrix (derivative of the forces on all atoms
when one atom is displaced in x, y, z directions) is plotted as a function of the num-
ber of orbitals in carbon fullerenes of different sizes. (Right panel) Performance of
the alternative algorithm described in the text [based on Eq. (22), blue circles], as
compared to the original implementation based on Eq. (20) (black diamonds). The
initialization (computing Ξ and Ω) is the most time-consuming step, although it
has to be performed only once, and can be used for all perturbations (each atomic
FIG. 11. Electron deformation density isosurfaces (blue positive and red negative) displacement). The new algorithm becomes more efficient for system sizes larger
for a proton displacing leftwards at a velocity of 1 a.u. in the bulk of a Ge crystal. than the threshold value (green dashed line).

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-13

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

∗
The solution of the Sternheimer equation and calculation of the +∂ρμν ∼ ∑ ciμ ∂ciν = ∑ Λ(l) ∗
νη ∑ ciμ ω̃l,i ciη
perturbed density matrix are the most demanding step. It requires i lη i

= ∑ Λ(l) (l) (l) (l) (l)

the perturbed coefficients of the electronic wavefunctions to be
νη Ωημ = ∑ Ξγν Δγη Ωημ ,
obtained, lη lηγ

∗ ∂Hαβ − εi ∂Sαβ (i)

∂ciμ = ∑ ∑ cjα [ ]ciβ cjμ = ∑ Aij cjμ , (20) where only the central term requires self-consistency and Ωαβ
j αβ εi − εj j ∗
= ∑j cjα ω̃i,j cjβ . Although computing the change in the density scales
c∗ Δi ciβ as O(Nb2 M), most of the computational cost is required in an initial-
where Aij = ∑αβ jαεi −ε
αβ
j
and Δiαβ = [∂Hαβ − εi ∂Sαβ ]. The change in ization step to obtain Ξ and Ω that are perturbation-independent,
the density matrix is then given by enabling the extraction of the whole dynamical matrix with O(Nb3 )
operations. A preliminary serial calculation for Cn fullerenes shows
all
∗ ∗ ∗ that the threshold system size for the new algorithm to become more
∂ρμν = ∑[ni ∂ciμ ciν + ni ciμ ∂ciν + ∂ni ciμ ciν ]
i efficient than the original implementation lies at around 650 atoms.
all all all
This value can be conveniently reduced by an efficient parallelization
= ∑ ni ∑[A∗ij cjμ
∗ ∗
ciν + ciμ ∗
Aij cjν ] + ∑ ∂ni ciμ ciν of the initialization step.
i j i

all unocc J. TRANSIESTA

= ∑ ni ∑ [A∗ij cjμ
∗ ∗
ciν + ciμ Aij cjν ] The transport code TRANSIESTA, initially developed by Brand-
i j
byge and co-workers,15 enables open-boundary condition calcula-
occ all tions by extending periodic regions with bulk electrodes. It is based
∗ ∗ ∗
− ∑ nj ciμ cjν ∑ cjα ∂Sαβ ciβ + ∑ ∂ni ciμ ciν , (21) on the non-equilibrium Green function formalism, which allows
ij αβ i
biased calculations. TRANSIESTA has been completely re-written and
and a similar expression applies to the change in the energy-density now uses advanced inversion algorithms, enables Ne ≥ 1 electrodes,
matrix. allows thermo-electric calculations, performing real-space calcula-
The change in the occupation of the electronic state can be tions (without k-points), and adds phonon transport calculations
computed from the change in its eigenenergy εi = ∑αβ ciα Δiαβ ciβ , and using the Hessian,97,98
it is relevant in metals for states close to the Fermi level. The Fermi The non-equilibrium Green function formalism can be sum-
level can also be shifted by the perturbation, and it can be deter- marized in the following equations, which are generalized for Ne ≥ 1
mined through the conservation of the number of electrons in the electrodes:
system, N e .
Obtaining ∂ρμν is the most computationally expensive part of ρ = ρeeq + ∑ Δee′ ≡ ρeneq , (23)
e′ ≠ e
the code. While the computation of ∂H μν basically has linear scaling
with the system size, the matrix Aij scales as Nb2 ⋅ M, where N b is
dkdε[Gk (ε) − G†k (ε)]nF,e (ε)e−ik⋅R ,
i
the number of basis functions and M is the maximum number of ρeeq ≡ (24)
2π ∬BZ
neighbor orbitals for any orbital in the system. Equation (20) then
requires Nb3 operations for each atomic perturbation, and the change
dkdε Ae′ ,k (ε)e−ik⋅R [nF,e′ (ε) − nF,e (ε)],
1
in the density matrix requires Nb2 ⋅ M loops. An alternative approach Δee′ ≡ (25)
that offers a better computational scaling for systems with a gap has 2π ∬BZ
∗
cjα cjβ
also been tested. If we define Ξiαβ = ∑j εi −εj
, we obtain
where ρeeq is the equilibrium density matrix for electrode e, G is the
retarded Green’s function matrix, and Δee′ is the correction to the
∂ciμ = ∑[∑ Ξiαμ Δiαβ ]ciβ = ∑ Λiμβ ciβ , (22) equilibrium part. The spectral function Ae = GΓe G† and carries elec-
β α β trons from the electrode e. Finally, nF,e is the Fermi function with the
chemical potential corresponding to electrode e. It is evident that the
where Λiμν = ∑η Ξiημ Δiην is a smooth function of εi and can be Fermi functions depend on the chemical potential and the electronic
described by Chebyshev’s expansion with a few selected energy temperature in the associated electrodes. By using different temper-
points and their corresponding weights, atures for each electrode, one can calculate thermoelectric effects
due to different reservoirs having separate electronic temperatures
(l) (l)
∂ciμ = ∑ ω̃l,i ∑ Λμβ ciβ = ∑ ω̃l,i ∑ Ξ(l)
αμ Δαβ ciβ . self-consistently.
l β l αβ We note that TRANSIESTA uses a multiple complex energy-
(l)
contour algorithm to more accurately describe the total density
Note that Ξαβ is perturbation-independent and could be com- ρ. It does this by weighing each ρeneq using a simple scheme97
puted only once and used for all the possible atomic displacements (Sec. III B). So far, few multi-electrode calculations have been per-
with a cost that scales as Nb2 ⋅ Nl , with N l being the number of formed, so the importance of the multiple contour algorithm is
Chebyshev’ polynomials. The change in the electronic density is then currently unknown.99–101 However, for the well-known 2-electrode
given by problem, it allows smoother convergence properties.15

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-14

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

In the latest TRANSIESTA, we implement three different inversion Wannier functions (MLWFs)107,108 and using them to compute the
algorithms: (i) a block-tri-diagonal (BTD) algorithm, (ii) a MUMPS advanced electronic properties of materials with high efficiency and
sparse algorithm, and (iii) a dense algorithm (LAPACK). The per- accuracy.
formance of these (speedup compared to TRANSIESTA 3.1) is sum- The Wannier functions can be considered as a unitary transfor-
marized in Fig. 13. Since the BTD algorithm is linear scaling for mation (more precisely, a Fourier transformation) of a set of Bloch
constant width, it can easily outperform the dense algorithm by a functions associated with a given manifold of bands. We can view
factor of 100. This performance gain is also important for the mem- the Bloch and Wannier functions as providing two different basis
ory footprint, enabling even larger systems. The BTD algorithm sets describing the same manifold of states associated with the elec-
favors long and narrow systems but uses less memory for all types of tron band manifold in question. The Wannier functions display a
systems. number of very interesting properties.109 Among them, we can enu-
A recent addition to the TRANSIESTA package is the use of real merate the following: (i) they are localized in real space, each of
space self-energy terms.98,102 These self-energies are semi-infinite in them concentrated around a given unit cell (see Fig. 14); (ii) Wan-
more than one direction and can thus be used as surrounding elec- nier functions centered on different cells are translational images of
trodes for, e.g., single defects in 2D or 3D structures or line defects. one another; (iii) they form an orthonormal basis set; and (iv) they
Real space self-energies are superior to BZ integrated quantities since span the same subspace of the Hilbert space as is spanned by the
they correctly describe the infinite bulk by leaving out image cou- Bloch functions from which they are constructed. Because of the
plings and also removing the need for k-point sampling. When gauge freedom in the definition of the phases of the Bloch func-
taking into account the complete procedure for a TRANSIESTA calcu- tions, the Wannier functions are not unique. However, the location
lation, the real space self-energies provide an increased throughput of their centers in the home unit cell is unique to within a lattice
since the SCF k-point sampling and the subsequent k-point sampled vector, i.e., they are gauge invariant.109 The high degree of arbi-
transport calculation are completely removed.98 trariness in the definition of the phases can be exploited to produce
Additionally, TBTRANS enables the calculations of user defined unitary transformation matrices between Bloch and Wannier func-
tight-binding models and also interfaces to phonon transport using tions in such a way that a localization functional that measures the
the Hessian matrix (program named PHTRANS). The phonon Green sum of the quadratic spreads of the Wannier functions in the home
function is similar to the electron one, unit cell around their centers is minimized.107 In a practical pro-
−1 cedure to construct Wannier functions, a set of localized functions
Gq (ω) = [(ω2 + iη2 )I − Dq − Σq (ω)] , (26) is used to generate an initial guess for the unitary transformations.
These localized functions should be roughly located on sites where
where D is the Hessian and ω is the phonon frequency. Finally,
Wannier functions are expected to be centered and have appro-
inelastic transport involving phonon-excitation can be treated with
priate angular character. In our implementation, we can directly
perturbation theory in a post-processing step with the INELASTICA
use the localized atomic orbitals of the basis or the hydrogenoid
package.103,104
localized functions (including hybrid orbitals), as suggested
in WANNIER90.
K. Wannierization The Wannier functions provide an exact tight-binding rep-
The interface between SIESTA and WANNIER90105,106 (version 3.0.0) resentation of the dispersion of the Bloch bands. This property
has been implemented, so the latter code can be called directly will be exploited to extract in an automatic and user blind way
from SIESTA as a library or used as a post-processing tool. WAN- the parameters required to run multiscale simulations, as described
NIER90 is an open-source code for generating maximally localized in Sec. III L.

FIG. 13. Performance characterization of TRANSIESTA using

a pristine graphene cell (24 atoms wide). Speedup for (a)
EGF and (b) NEGF calculations of pristine graphene com-
pared against the dense implementation. The BTD method
exhibits more than 40 times the speed of the LAPACK
implementation for the largest size. MUMPS gains speed
after 5000 orbitals.

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-15

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

Currently, WANNIER90 can be used as a post-processing tool or accurate force-field114 that allows for fast evaluation. The E(1) and
it can be directly called from SIESTA in a library mode. Within E(2) terms account for the changes in the electronic structure that
this latter approach, the unitary matrices that transform the Bloch are represented by geometry-dependent Wannier functions. Under
states into Wannier functions are directly accessible in SIESTA, allow- this basis, E(1) becomes a tight-binding model, while E(2) represents
ing a clear and straightforward interconnection between the two electron–electron interactions.
alternatives to span the Hilbert space. Besides, the use of Wan- The interconnection between the first (SIESTA) and second
nier functions opens the door to a wide range of potential appli- (SCALE-UP) principles simulations is carried out through a python
cations. Already implemented in SIESTA is the possibility of per- script, MODELMAKER. Taking a few cutoff distances, MODELMAKER
forming SCF convergence under the constraint of a rigid shift on is able to produce a model’s terms and automatically carry out
the energy associated with a given Wannier function to be used DFT simulations with SIESTA to determine the force field, a Wan-
to calculate electron–electron interactions for multiscale simula- nier Hamiltonian to represent the bands, electron-lattice terms to
tions, as detailed in Sec. III L. The interface with the self-consistent account for how the bands change with geometry, and electron–
dynamical mean field theory DMFTWDFT code110 using MLWF has electron interactions to describe, for example, magnetism.
been already implemented.111 In addition, alternative approaches While, so far, few publications with SPDFT methods include
to compute the exact Hartree–Fock exchange in extended insulat- explicit treatment of electronic degrees of freedom, the lattice part
ing systems with a linear scaling computational cost using MLWFs has successfully been used in several applications. One of the main
have been proposed, being another interesting research line for the fields of research has been thermal conductivity in perovskites. In
future.112 particular, it was employed to study the electrophononic coupling
in SrTiO3 115 and PbTiO3 116 and the proposal of a thermal switch in
PbTiO3 .117 It has also been used to study the competition between
L. Multiscale methods various ferroelectric domain structures in PbTiO3 /SrTiO3 super-
Density functional theory can be used as the basis for parame- lattices as a function of strain.118 As a result, it was found that
terized multiscale methods that can be used to carry out simulations tensile strains lead to the appearance of chiral ferroelectric vor-
including tens or even hundreds of thousands of atoms.113 First- tices, while ferroelectric skyrmions were predicted and experimen-
principles methods are used to produce detailed models that are tally observed for more compressive strain values.118 The calculated
subsequently used to predict properties that require large-scale sim- dielectric properties of these superlattices119 are in very good agree-
ulations. The models are created for specific materials, and their ment with measured values and show very large electric suscep-
accuracy can be systematically improved to converge toward DFT tibility consistent with regions of negative, static electric permit-
precision. Given the dependence on first-principles, we refer to these tivity situated at the core of the vortices and the PbTiO3 /SrTiO3
methods as second-principles DFT (SPDFT) and run them on an interfaces.
independent code called SCALE-UP.113
SPDFT is based on a division of the total electronic density, M. Scripting and integration in external frameworks
n(⃗r), into reference [n0 (⃗r)] and deformation [δn(⃗r)] contributions,
An ongoing trend in many areas of computational science is
to move away from rigid and monolithic codes, favoring instead
n(⃗r) = n0 (⃗r) + δn(⃗r), (27)
a more flexible approach in which the internal functionality of a
program is somehow exposed to the outside world. If done in a
where δn(⃗r) is considered as a small perturbation with respect proper and well-documented way, this can serve to enhance the
to n0 (⃗r) that, in non-magnetic cases, represents the ground state interoperability of codes with different functionalities, playing to
of the system.113 This division is then used113 to expand the the relative strengths of each, and/or to implement new function-
DFT energy with δn finding that the zeroth order term, E(0) , alities by combining the available basic blocks. In SIESTA, we have
corresponds to the full DFT energy for the reference density. followed two different but complementary routes to these ends: the
The corrections to this reference energy only depend on δn development of an internal scripting framework based on the Lua
(and parametrically on n0 ) which, given its smallness, can be language, which enables new functionality without code recompi-
efficiently calculated leading to a fast and accurate approxima- lation, and the implementation of a formal interface to the AiiDA
tion of the full DFT energy. The expansion is usually taken to platform.
second-order,

E ≈ E(0) + E(1) + E(2) + ⋯,

1. Lua interface
(28)
Lua120 is an easy-to-learn and fast scripting language built for
resulting in a stationary problem that is equivalent to Hartree–Fock embedding. It is very lightweight (its memory footprint is less than
with the important distinction that the interactions are screened by 300 kB) and provides very simple ways to interface to the data struc-
the exchange–correlation potential. In order to keep δn small, the tures and routines of a host program. A Lua script, interpreted by
application of the method is restricted to problems where atomic the Lua interpreter embedded in the program, can then control the
bonds are not created or destroyed, i.e., to processes that display an flow of execution and the data. Different user-level scripts can imple-
invariant bond topology. ment new functionalities, without recompilation of the host code.
The E(0) term represents the exact DFT energy for the refer- The strategy we have followed in SIESTA is based on handling control
ence density. We represent it for a variety of geometries with an to the Lua interpreter at specific relevant points in the program flow

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-16

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

FIG. 14. Maximally localized Wannier functions (MLWFs) for graphene. Panel (a) displays the character of σ-bonded combinations of sp2 hybrids. Panel (b) displays the π
character of the bands with weight on the pz orbitals. Isosurfaces of different colors correspond to two opposite values for the amplitudes of the real-valued MLWFs. Yellow
spheres represent the position of the C atoms, while smaller blue spheres mark the center of the bonding.

(e.g., at the beginning of a geometry step and at the end of a operations of preparing the input files for a calculation using
SCF step). Lua scripts implement handlers appropriate to the point AiiDA-specific input objects, parsing the results and generating
they want to hook into and can request access to specific data the AiiDA output objects. The AiiDA data are stored in a graph
structures. For example, a script intended to implement a better database that keeps a permanent record of the inputs and outputs
SCF mixing algorithm would be executed after every SCF step, of the calculation and is fully searchable for, e.g., data analytics
inspecting the convergence data and changing mixing parame- purposes.
ters or schemes, as appropriate. As another example, convergence AiiDA also provides robust support for the creation of work-
checks over mesh-cutoffs and k-point sampling can be performed flows that incorporate all the necessary steps in the calculation of
automatically. potentially complex properties, including the proper heuristics and
The above mixing scenario exemplifies an important area of fail-safe features. The aiida-siesta package provides a base work-
usefulness of the approach: the prototyping in Lua (followed even- flow and a few workflows for standard materials properties, such as
tually by a full implementation) of new ideas and algorithms. We band structures. Figure 15 shows the execution graph of a work-
have implemented a number of custom molecular dynamics (MD) flow designed to generate a synthetic scanning tunneling microscopy
modes, geometry relaxation algorithms, and advanced optimization (STM) image from a given structure. Work is ongoing to implement
schemes in a pure Lua library FLOS.121 The code in the library can more complex ones.
be re-used or taken as a starting point for other implementations by In addition to an interface to the computational capabilities of
users. These user-level scripts can, in turn, be shared, opening the the SIESTA code via the plugin and workflows, the aiida-siesta
way to the development of new functionality with faster turnaround package also provides an implementation of basic objects represent-
that the traditional approach that needs a careful integration into the ing pseudopotential files, notably one for PSML. Families of pseu-
program’s code base. dopotentials can be uploaded to an AiiDA database and shared via
As a specific demonstration of the power of the Lua embedding, the provided mechanisms for data export and import, facilitating the
we have developed a number of variations of the nudged-elastic band interoperability of different codes.
(NEB) method122,123 for transition-state search. Previously proposed
implementations in SIESTA involved significant, hard to maintain
code changes, and did not make into the mainstream version. With N. Utilities for post-processing and supplementary
Lua, we have been able to implement, non-intrusively, not only the features
standard algorithm, but a Double Nudged Elastic Band (DNEB)124
variation, and also another version that treats atomic coordinates SIESTA offers several features beyond the core functionality of
and lattice variables on an equal footing (the variable-cell NEB or solving the electronic structure problem and performing optional
VC-NEB method125 ). geometry relaxations and molecular dynamics runs. It is worth not-
The integration of Lua functionality in SIESTA has been made ing, in particular, that the atomic character of the basis set enables
possible by the development of an intermediate layer, FLOOK126 (for the use of a very intuitive suite of analysis tools, which take advan-
“fortran-Lua-hook”), which provides wrappers for access to Fortran tage of the fact that most of the concepts relating to chemical
data structures and subroutines. bonding use the language of atomic orbitals.
The (partial) density of states, atomic and orbital popula-
tions, and other useful outputs can be obtained directly from
2. AiiDA plugins and workflows the program. The SIESTA distribution also includes several tools
The AiiDA framework127–129 provides support for high- in the Util directory for band-structure and wavefunction plot-
throughput computations in materials science, keeping full prove- ting, bonding analysis, etc. Beyond these, special tool packages
nance of the calculations and facilitating data handling and sharing. that implement a specific feature that extends the functionality
The framework is open-source, written in Python, and designed of the main program or that provide extra options for visualiza-
to support arbitrary codes via a plugin interface. A plugin for tion or post-processing, in general, are available in alternate dis-
SIESTA has been implemented and is distributed as the open- tribution points. We describe in what follows the most relevant
source package aiida-siesta.130 The plugin provides the basic developments.

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-17

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

FIG. 15. Automatically generated graph for the execution of an AiiDA workflow for the simulation of STM images.

1. Updates to core utilities 2. SISL

A number of improvements, enhancements, and additions have SISL is a Python toolbox that was initially conceived to handle
been made to the core utilities shipped with the SIESTA distribution. and manipulate the SIESTA/TRANSIESTA output.102 It has since been
There is now a “fat-bands” feature by which bands can be deco- extended to support other DFT codes with the aim of offering
rated with information about the relative weight of the given orbitals equivalent operations for them.
in each state. The wavefunction-related analysis tools have been By reading the LCAO outputs from SIESTA, one can post process
extended to the non-collinear and spin–orbit cases. This includes the Hamiltonian and calculate, e.g., Brillouin zone integrated DOS,
the COOP/COHP bonding analysis, band-structures, and a new wavefunctions expanded on grids, eigenvalues, and band velocities.
tool for spin-texture calculation. There have also been improve- SISL can process nearly all the SIESTA output files. In particular, it is also
ments to band-structure plotting utilities and to the visualization of able to post process the data on the real-space grid. Its command
charge densities, potentials, and other magnitudes represented in a line interface allows data format changes, e.g., conversion of SIESTA
real-space grid. XV files to xyz/xsf files or SIESTA binary grid data (VH, VT, etc.) to
A band unfolding utility has been added. Based on the Fourier cube/xsf files.
decomposition of the Bloch wavefunctions, it allows us to perform As it can process density matrices from SIESTA, one can also
a “full unfolding” even for non-periodic systems (e.g., liquids) cal- use SISL to prepare an input electronic structure for new calculations,
culated with a large simulation cell. By refolding the fully unfolded which may be helpful to reduce initial SCF steps.
bands from the reciprocal supercell of a perturbed or defective crys- SISL also allows the creation of custom tight-binding models
tal to the reciprocal unit cell of the primitive crystal, one recovers the (both orthogonal and non-orthogonal), and since it extracts the DFT
conventional “unfolded” bands.131 Hamiltonian matrix, one can manipulate the Hamiltonian to retain

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-18

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

certain band-structure features and thus perform large-scale simu- can be found in Ref. 136. The first type of projection, if done for a
lations.132 This allows calculating far-field currents using reduced sequence of q values, helps to reveal “phonon dispersions,” obviously
basis-sets with a very little loss of accuracy. blurred by the broken periodicity, also making distinction between
The Atomic Simulation Environment (ASE)133 and SISL have a transversal and longitudinal modes (see Ref. 137 for an example of
certain degree of overlap in terms of geometry handling function- use). To make the trends more pronounced, the supercell needs to
ality. One can easily convert to and from ASE objects in SISL, thus be sufficiently long in the direction concerned (see, e.g., Fig. 16). The
allowing seamless interaction. simplest case, a projection onto a single q = 0 value, may also be of
interest, since it enhances the modes that are expected to dominate
the infrared or Raman spectra and thus facilitates their comparison
3. Other post-processing and visualization utilities
with the experiment.
The body of utilities contributed by non-core developers and The symmetry projection may help to isolate in a possibly
other SIESTA users has continued to expand. In particular, we feature complex spectrum those modes which are expected to dominate
in this section two suites of utilities: one dealing with alternate visu- according to a given selection rule, again in view of their verifi-
alization tools for some SIESTA results and another one specifically cation against the experiments. The group-symmetry information
dealing with lattice dynamics. needed for the projections is available, e.g., from the Bilbao Crys-
For structures, the xv2xsf and xv2vesta converters process tallographic Server,138 and the technical details are explained in the
data from the SIESTA .XV file into the native formats of XCrySDen134 documentation included in the tools.
and VESTA,135 respectively. Each of these two codes offers many The vibent tool performs a straightforward calculation (see,
options of graphical representation of structures, adding trans- e.g., Sec. II.C in Ref. 139 or Sec. 5.3 in Ref. 140) of temperature-
lations, clipping fragments, etc. Three-dimensional spatial func- dependent vibration contributions to the free energy and entropy
tions (e.g., charge density and local density of states integrated (see Fig. 17 as an example). The necessary input information is
throughout the chosen energy range) are computed by SIESTA on the vibration spectrum, originating from the Vibra frozen phonon
a real-space grid. Tools are provided for interpolating the data calculation on a sufficiently large system.
from the SIESTA output grid (fixed by the unit cell dimensions and The velcf tool calculates the velocity autocorrelation func-
the MeshCutoff parameter) onto an arbitrarily cut (and possibly tion and its Fourier transform from a (presumably sufficiently long)
rotated or resampled) parallelepipedic box. XCrysDen provides a molecular dynamics (MD) history, recorded in the .MD or .ANI file.
number of display options, including contour lines over grid planes This technique141 can be used to obtain phonon frequencies and was
or isosurfaces. A special feature available in XCrySDen is plotting applied along with a SIESTA calculation in Ref. 142. An example of
the Fermi surfaces. A special script, eig2bxsf, serves to analyze such simulation (1000 MD steps at 600 K) is shown in Fig. 18 in
the list of k-points handled by SIESTA, expanding it onto a regular comparison with frozen phonon results, revealing similarities of the
sequence and writing the respective band energies in the necessary spectra obtained.
format.
The tools concerning the lattice dynamics have been devel- 4. Optical properties of finite systems: Linear
oped having in mind the Γ phonons calculated for a large enough response TDDFT starting from SIESTA orbitals
supercell, that is a typical case in a simulation of molecular crystals
or disordered substitutional alloys. For visualization, vib2xsf and The SIESTA package offers at least two ways of obtaining the
vib2vesta place arrows at the atoms according to the vibration pat- optical properties of finite systems. The first way uses real-time
tern stored in the eigenvectors file (.vectors), produced by the core TD-DFT propagation by applying an external electric field with a
Vibra utility, and can also be used to make animations (sequences simple time dependence (e.g., a Heaviside step-function).74 The sec-
of snapshots) of the selected vibration modes. Both vib2xsf and ond way is by computing the non-interacting dielectric function.1,2
vib2vesta tools allow the selection of a part of the system to be Both methods are implemented in SIESTA and can be employed with-
exposed. out any external tools. However, they are limited in different aspects.
The phdos tool is designed for analyzing the zone-center vibra- The non-interacting dielectric function often underestimates the
tion results. As the system is supposed to be large (e.g., a supercell HOMO–LUMO gaps and calls for the use of the phenomenolog-
chosen for a periodic crystal), the (artificially broadened, for con- ical scissor-shift operator. Real-time propagation makes cumber-
venience) discrete spectrum may serve as a fair approximation to some the analysis of the optical response properties in the frequency
the total density of modes, and if weighted with (squared) compo- domain. Furthermore, the frequency resolution scales with the dura-
nents of eigenvectors at different atoms, it provides a decomposition tion of the real-time simulation. Thus, accurate spectra require long
into contributions of different atoms in the total density of vibration simulations.
modes. Fortunately, there are two efficient implementations of linear-
A more sophisticated option is the projection of different eigen- response TDDFT that use the Kohn–Sham orbitals from SIESTA as
vectors according to various criteria. The typical system under study a starting point and are available for the open-source commu-
is a supercell in which, e.g., an alloying, or some kind of deformation, nity.143,144 In both packages, the linear density response δn(r, ω) is
breaks the underlying perfect periodicity. Still, some trends related obtained directly in the frequency domain, which makes the analysis
to the latter can be revealed by appropriate projections. The two of derived properties straightforward. However, there are differences
obvious cases are the projections onto (1) q-vectors of the under- between both implementations on the construction of the auxiliary
lying lattice and (2) irreducible representations of the space group of basis necessary to expand the orbital products. These differences can
the underlying lattice; the corresponding formulas and some results severely affect the computational cost of the calculation.

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-19

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

FIG. 16. Left panel: a 192-at. quasirandom supercell repre-

sentative for the Be1/3 Zn2/3 Se solid solution and right panel:
density of modes within the frequency range of Zn–Se vibra-
tions, extracted with phdos and projected onto different val-
ues of qz and different polarizations, parallel (labeled LO)
and perpendicular (TO) to q. These results were partially
shown in Fig. 4 of Ref. 137 and discussed in that work.

The linear-response TDDFT is built on the concept of the χ 0 (r, r′ , ω)145 with a Dyson equation,
induced electronic density δn(r, ω) in response to a small per-
turbation of the external potential δV ext (r, ω). The integral oper- χ(ω) = χ0 (ω) + χ0 (ω)Kχ(ω), (29)
ator connecting δn(r, ω) to δV ext (r, ω) is the interacting density
response function χ(r, r′ , ω). By virtue of the KS equations, χ(r, r′ , ω) where the interaction kernel K(r, r′ ) contains the bare Coulomb
can be connected to the non-interacting density response function interaction and the so-called exchange and correlation kernel K xc ,

FIG. 17. Vibration properties of

Cu2 ZnSnS4 (CZTS) with substitutional
impurities used in Ref. 139. Left panel:
densities of modes (extracted with
phdos) and right panel: vibration contri-
butions to the free energy and entropy
(calculated with vibent). Adapted from
Figs. 5.3 and 5.4 of Ref. 140.

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-20

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

FIG. 18. Vibration properties of “Ni4 ” molecular magnet [Mo12 O30 (μ2 -OH)10 H2 {Ni(H2 O)3 }4 ]⋅14H2 O. Left panel: density of modes from frozen phonon calculation and right
panel: velocity autocorrelation function, its Fourier transform, and the resulting density of vibration modes.

which is a known operator for simple functionals such as LDA and and rather potent. It is capable of computing the optical proper-
GGA. The non-interacting response function χ 0 (r, r′ , ω) can be ties of compact metallic objects containing up to several hundreds of
expressed as a sum over electron–hole excitations within the basis atoms.146,149,150 For example, we were able to track down the differ-
formed by the KS orbitals Ψn (r),144–146 ent size-dependences of the plasmon resonance in sodium and silver
clusters due to the screening effect of silver d-orbitals in the latter
Ψn (r)Ψm (r)Ψm (r′ )Ψn (r′ ) case.151 In those calculations, using an optimized version that incor-
χ0 (r, r′ , ω) = ∑( fn − fm ) , (30) porates some additional memory-saving features not present in the
nm ω − Em + En
currently distributed version of PySCF-NAO, icosahedral silver and
sodium clusters containing up to 5043 atoms were studied.
where fn are occupations of the KS orbitals and En are their energies. In Figs. 19 and 20 we show the photo-absorption cross sec-
The optical polarizability tensor α(ω) is related to the induced tions of a series of compact silver clusters146 and the real part
density by α(ω) = ∫rδn(r, ω)dr or, alternatively, of the induced density change in the cluster Ag147 close to its
surface-plasmon frequency (3.4 eV), respectively.
α(ω) = ∬ rχ0 (r, r′ , ω)δVs (r′ , ω) drdr′ , (31)
5. Thermal transport by the AEMD method
where due to Eq. (29) and using the dipole approximation for The approach to the equilibrium molecular dynamics (AEMD)
the electron–photon coupling, the screened effective perturbation method152 has been implemented to obtain the thermal
δV s (r′ , ω) satisfies the linear integral equation,

(I − Kχ0 (ω))δVs (ω) = r. (32)

The efficiency of the methods presented in Refs. 143 and 144

comes from solving iteratively Eq. (32) for δV s (ω) instead of using
standard matrix inversion to obtain χ(ω) from Eq. (29). Once δV s (ω)
is known, Eqs. (30) and (31) allow the computation of the optical
properties of the system. Furthermore, it is also possible to per-
form different types of analyses. For example, it is easy to partition
the polarizability tensor α(ω) in terms of electron–hole contribu-
tions146,147 due to the existence of the sum over the electron–hole
pairs in Eq. (30). Similarly, one can achieve other types of Mulliken-
like analyses144,146,148 of the optical polarizability tensor α(ω) or the
induced density δn(r, ω). FIG. 19. The absorption cross sections of silver clusters of icosahedral shape. One
The Python implementation of linear response TDDFT in the can recognize sharp surface-plasmon resonances around 3–4 eV and a broad
resonance at 6–7 eV.
PySCF-NAO package as described in Ref. 144 is convenient to use

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-21

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

differences in the energy of the crystal with a “core-hole” in different

sites. For this, a special pseudopotential with a missing core electron
has to be generated, and a full SIESTA calculation is needed for each
different site.
The implemented methodology has been used to study, for
example, the shifts induced by hydrogen bonding in organic
molecules.155

O. Software-engineering advances and partnerships

The traditional development model for scientific codes in
academic settings has been typically based on multiple contribu-
tions with various levels of programming competence and with
very little time to plan ahead in the face of pressing scientific
demands. SIESTA has been no exception and has grown in features
and complexity over the years. It is very important to keep com-
plexity under control or else a project becomes un-maintainable
FIG. 20. The isosurfaces of density change Re(δn(r, ω)) of the Ag147 cluster close and cannot survive. It is not simple, however, to balance the
to the frequency of the surface-plasmon resonance of the cluster (3.4 eV).
need of incorporation of new features and the need to increase
the computing performance in a landscape of constantly evolv-
ing hardware and programming models. One essential route is
conductivity. In the first stage of the method, the system is decom- modularization, which allows the separation of concerns at var-
posed into two different regions, each one equilibrated to a differ- ious levels. In the context of a code such as SIESTA, this means
ent initial temperature (canonical run with Bose or Anneal MD). that the scientific ideas and algorithms should be handled at a
Then, a microcanonical run (Verlet) is carried out for the whole high level, calling on lower-level modules for specific functional-
system, and the average temperature of each subsystem is moni- ity (domain-specific libraries, mathematical libraries, communica-
tored. This temperature transient regime is then used to extract the tion protocols, etc.). These lower-level modules can hopefully be
thermal conductivity from the exact solution of the heat transport re-used by different codes and, most importantly, can be focused
equation.153 on by highly skilled programmers for optimization on relevant
architectures.
Another important method of taming complexity involves
6. Core-level shifts the streamlining of the data structures of the code. This is
Core-level shifts can serve to analyze changes in the local an ongoing process (see Sec. V) but has already taken a very
and chemical environment of atoms of a given species. Density- significant step by the introduction of reference-counted data
functional-theory calculations have proved to be quite useful in structures. They build on a well-known and not particularly
complementing the experimental information, which is sometimes advanced technique of memory-handling,156 but in SIESTA they
hard to interpret. Two schemes have been implemented in SIESTA have enabled a much simpler bookkeeping of the data structures
for the calculation of core-level shifts within a pseudopotential needed for a richer control of molecular-mechanics (MM) and SCF
approach.154 iterations.
In the so-called initial-state approximation, the electronic Regarding performance-oriented developments, in the recent
relaxation in the presence of the core hole is neglected, and the past we have implemented a mixed MPI/OpenMP program-
photo-electron’s binding energy is directly related to the eigenvalue ming model, which allows, for suitable systems, to better bal-
of the core level. A pseudopotential calculation obviously cannot ance arithmetic intensity and communications needs. The deploy-
compute the latter, but differences in core eigenvalues in different ment of this model is more advanced in the TranSiesta mod-
environments can be estimated by the changes in the expectation ule, and significant speedups have been obtained for large
value of the crystal potential using the core state’s atomic wavefunc- systems.
tions ψnlm at different sites. These can be extracted from the matrix Some of the above software-engineering developments have
elements, been enabled and strengthened by the participation of SIESTA in a
number of international partnerships, notably the MaX (Materials
at the eXascale) EU center of excellence157 and the Electronic Struc-
V mm = ∫ d3 r (ψnlm (⃗r − ⃗τ ))∗ V(⃗r)ψnlm (⃗r − ⃗τ )
′ ′
(33)
ture Library initiative.158 The “separation of concerns” described
above in the context of modularization is an example of the so-called
with a further step of averaging to remove the splittings stemming “open-innovation” paradigm, at the foundation of the ESL strategy
from the loss of spherical symmetry. for code reusability, and is also a cornerstone of MaX’s efforts to
In the final-state approximation, the relaxation is explicitly achieve exascale-readiness for its flagship materials science codes
taken into account, and the experimental shifts (measured via (with SIESTA among them): performance-enhancement efforts are to
the kinetic energy of an existing electron) are correlated with the be focused on relevant domain-specific modules.

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-22

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

A number of modules from SIESTA have been turned into combined with advanced calculations can provide new insights into
stand-alone libraries that now feature in the ESL: libGridXC for transport properties at the nanoscale.
exchange and correlation calculations, libPSML as a handler of
PSML files, xmlf90 for general purpose handling of XML files, etc.
B. Novel topological phases in ferroelectric materials
Conversely, SIESTA uses some of the libraries offered by the ESL,
notably the ELSI library of electronic structure solvers mentioned in In material systems with several interacting degrees of freedom
Sec. III G 2, whose development, including its API design and inter- (such as spin, charge, and lattice distortions), the complex inter-
nal data organization, has been, in turn, influenced by contribu- play between these factors can give rise to exotic phases. Prototyp-
tions and feedback from the SIESTA project, among others. There are ical examples are the superlattices of alternating lead titanate and
also plans to incorporate the PSolver library159 for the solution of strontium titanate layers. Simulations on such PbTiO3 /SrTiO3 het-
the Poisson problem, a contribution to the ESL from the BigDFT erostructures, consisting of n unit cells of PbTiO3 and n unit cells
project. of SrTiO3 stacked along the [001] direction, were carried out with
We should mention that the renewed dynamism of SIESTA devel- SIESTA. As a function of the periodicity, the superlattices undergo a
opment and the advances made possible by the interaction with phase transition from a monodomain configuration (small period-
community initiatives are both a blessing and a challenge. It is icity, n ≲ 3–4) with a normal component of the polarization that
non-trivial, for example, to handle the building process of a code is preserved throughout the structure, to a multidomain configu-
that relies on a number of different external libraries, programming ration (large periodicity, n ≳ 3–4) with alternating up and down
models, and special features such as the embedded Lua interpreter. domains.160 In order to further reduce the electrostatic energy costs,
Fortunately, as will be discussed in Sec. V, these are issues that are the local dipoles within the PbTiO3 layer continuously rotate form-
being addressed in wider contexts, and SIESTA is well placed to take ing a sequence of clock-wise/counter-clockwise array of vortices
advantage of it. along the [100] direction. The theoretical predictions, done with
SIESTA161 after the relaxation of supercells of up to 1000 atoms, were
experimentally confirmed five years later by atomic-scale mapping
IV. APPLICATIONS of the polar atomic displacements by scanning transmission electron
We present here a few demonstrated applications that illustrate microscopy162 (Fig. 22). Moreover, the appearance of an axial com-
the capabilities of SIESTA in breadth, efficiency, and accuracy. ponent of the polarization pointing in the direction of the vortices
makes the systems chiral and optically active, as recently confirmed
by circular dichroism experiments.163
A. 4 terminal NEGF on germanium surface
Breakthrough simulations using the new multi-terminal imple-
mentation in TRANSIESTA were fundamental to elucidate the elec- C. 1D and 2D systems
tronic transport mechanism on a novel and complex exper- SIESTA is particularly well suited to study low dimensional
iment.100 For the first time, two-probe scanning tunneling nanostructures, such as 1D and 2D systems, where a large vacuum
microscopy/spectroscopy (STM/STS) with probes operating in tun- region is needed within the simulation cell. When, in addition, a
neling conditions over the same atomic-scale system was used to large number of atoms are required to study particular physical
extract detailed information on in-plane electronic transport. The effects, SIESTA could excel with respect to other methods. There is
addressed system was the reconstructed (001) surface of germanium, an extensive literature on simulations of graphene and other exfoli-
where electrons injected from one STM tip at a position determined ated materials, where the properties of point defects, edges, or grain
with atomic precision were collected at the same Ge dimer row boundaries are of much relevance. We list a few examples the mag-
at a distance as short as 30 nm. The experiment was theoretically netic properties of impurities164,165 and edges,166 electronic proper-
modeled by a system composed of a 12-layer Ge(001)-c(4 × 2) slab ties, including transport characteristics, in grain boundaries,167,168
contacted by Au tips oriented along the (100) direction (Fig. 21). ribbons,169 nanoporous graphene,170 large graphene flakes,68,171 or
On this self-consistent 4-terminal treatment, two Ge electrodes were the effect of substrates.172 Other materials, such as mono- and multi-
connected at each slab termination and other two at the Au model layered dichalcogenides173,174 or phosphorene,175,176 are also being
tips. The whole system was defined by 4924 atoms (36 442 atomic widely studied, including optical properties in nanoflakes with up to
orbitals) in a supercell of dimensions ∼32 × 160 × 80 Å3 , where a few thousand atoms.177
five different tip-to-sample distances were considered. Besides the
large dimensions of the system, another important challenge of such
simulation was the level alignment between the metallic and semi- 1. CDWs
conducting leads and the scattering region for which a method had A number of recent studies on charge density waves (CDW)
to be devised. A remarkable agreement was found between the calcu- in low dimensional materials illustrate the impressive accuracy that
lated transmission function and the experimental transconductance can be obtained with SIESTA for systems with very subtle electronic
spectra, allowing the identification and assignment of the observed structures.178 For example, in 2H–NbSe2 , SIESTA calculations were
resonances to transport channels existing along the surface Ge dimer able to predict the existence of six different atomic structures within
rows. Moreover, the simulations elucidated the transport direction- a narrow energy range of a few meV, all of them compatible with
ality of the injected hot electrons, revealing a transition from the the experimental 3 × 3 CDW modulation. Careful analysis of the-
2D to quasi-1D coherent transport regime as a function of the car- oretical and experimental STM images for different bias potentials
rier’s energy. This work shows that complex experimental setups allowed us to identify two of these structures that can coexist in

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-23

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

FIG. 21. First-principles transport simulations for the two-probe experiments. Representation of the four-terminal setup. The electrode regions are highlighted by blue boxes,
two of them located at each Ge(001)-c(4 × 2) slab termination (leads: left and right) and the other two at each Au model tip (leads: tip1 and tip2). The 50 Ge atoms closest to
each tip were allowed to fully relax. Adapted from Ref. 100.

the same image.179 In a different work,180 the temperature depen- one can obtain the coherence length of the fluctuating 1D electron–
dency of the electronic Lindhard response function in blue bronze hole pair (that determines the length scale of the experimental intra-
K0.3 MoO3 was studied. This system has a rather complex mono- chain CDW correlations) and the intrachain modulation of the
clinic structure with 20 formula units per unit cell where MoO6 response (that determines the shape of the Kohn anomaly mea-
octahedra form chains along one direction (b-axis). The Lindhard sured in experiments), providing, for the first time, quantitative evi-
function shows well decoupled sharp responses that correspond to dence of the weak electron–phonon coupling scenario for the Peierls
intra- and interband Fermi surface nesting. By fitting these peaks, transition.

FIG. 22. Panel (a): local polarization profile of polydomain structures in (PbTiO3 )n /(SrTiO3 )n with n = 6 obtained from an atomic relaxation with SIESTA. The PbTiO3 and SrTiO3
are depicted as gray and white regions, respectively. Clockwise and counterclockwise vortices within the PbTiO3 are clearly visible. Red dashed square in the SrTiO3 layers
mark the position where antivortices are formed. Reprinted with permission from Aguado-Puente and Junquera, Phys. Rev. B 85, 184105 (2012). Copyright 2012 American
Physical Society. Panel (b): experimental observation of vortex–antivortex structures in a cross-sectional high-resolution scanning transmission electron microscopy image
with an overlay of the polar displacement vectors for a (SrTiO3 )10 /(PbTiO3 )10 superlattice, showing that an array of vortex–antivortex pairs is present in each PbTiO3 layer.
Courtesy of R. Ramesh, adapted from Ref. 162.

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-24

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

FIG. 23. Color coded electrostatic poten-

tial on a plane cutting along the main axis
of the G. Sulfurreducens pilin molecule
in wet conditions. A perspective ball ren-
dering of the atomic structure of the pro-
tein is superposed. For the meaning and
details of this figure, see Ref. 184 (Figure
courtesy of Gustavo T. Feliciano).

D. SIESTA in biology: Pilin proteins as conductors like to refer the reader to a sample of recent reviews in various
SIESTA’s efficiency and the clear bandgaps of biomolecules, in fields in which the program is featured. These cover biological sci-
general, have made molecular biology a very suitable field for SIESTA ences185 (including interaction between organic and inorganic mate-
since the beginning181 and have stimulated the targeted develop- rials186,187 ), geology and materials under high-pressure,188 isotopic
ments of the code for the field, such as the hybrid Quantum Mechan- fractionation predictions for Martian geochemistry,189 the engineer-
ics/Molecular Mechanics approach (QM/MM).182,183 An interesting ing of typical core structural materials used in nuclear reactors,190
illustration of its suitability in an all-quantum biological problem or even in astrophysical and atmospheric systems.191 The reactivity
is the study of the electrostatics around the pilin protein in aque- of metallic nanoparticles for catalysis was studied by Viñes, Gomes,
ous solution.184 The pilin considered here is the main protein in the and Illas,192 and the role of SIESTA in the computation of the kinetic
pili (external filaments) of the Geobacter sulfurreducens bacterium, and dynamics of the catalytic reaction at surfaces (including adsorp-
which have been shown to be able to transmit electronic current, tion and desorption of reactants or products) was explored in Chap.
allowing the microbe to feed by remote redox reactions on ferrous 8 of Ref. 193 by Catapan and co-workers.
mineral particles in the soil. As a nanowire designed by natural
evolution, understanding the mechanism for charge transport is of
obvious interest. V. FUTURE EVOLUTION
Peculiar to this protein is the fact that its main alpha helix, Work on enhancing SIESTA’s capabilities, performance, and
the main feature of this elongated protein, is singly oriented, that robustness is continuing, driven by a good number of developers
is, there is no back alpha helix (as in a common hairpin configura- and collaborators. A mature and flexible development platform and
tion) that would counter the polarization of the single alpha helix. practices are essential to keep it productive. Our recent platform
In an alpha helix, all peptide-bond dipoles point in the same direc- changes have forced developers to shift workflows twice in the past
tion along the axis of the helix, which, in solid-state parlance, rep- four years. Through the changes, we have learned a lot but also
resents a polarization with clear electrostatic implications. Indeed, spent a significant amount of time on ensuring SIESTA’s continuous
a DFT calculation of the molecule in vacuum shows a well-defined development. At the current state, we believe we have stabilized the
electrostatic potential ramp along the protein, which tends to close development platform on GitLab, while we will add more integrated
the effective bandgap. The question is then, how does an aqueous development features in the coming years, e.g., continuous integra-
environment affect this depolarizing field. tion (CI) and source code checks. Using CI will also enable easier
Long molecular mechanics (MM) simulations were performed code-style checks to conform to coding standards. We hope that
for the protein in a suitable solution of NaCl at a concentration our open-platform initiative will keep external contributions coming
of 0.1M. The protein’s residues had charge states corresponding to into the program.
pH = 7, and the MM field was validated with SIESTA calculations in Our basic development plans include refactoring, apparently
vacuum (944-atom dynamic relaxation in a 104.43 Å3 box). The wet unexciting but essential to streamline the code base to enable fur-
system contained 4580 atoms, and the statistical average of the elec- ther implementations. In addition, we foresee a change in the release
trostatic potential around the molecule (see Fig. 23) was obtained model, moving away from coexisting long-lived release branches
from a sample of full SIESTA calculations of statistically independent whose maintenance takes up a lot of time and offering instead more
snapshots, taken every 50 ps during the last 0.5 ns of the simulation. frequent and short-maintenance releases.
Figure 23 shows how the aqueous environment kills the quite We plan to exploit the idea of modularization, continuing the
homogeneous potential ramp along the protein axis that appears in abstraction of relevant reusable pieces, but also dealing with a higher
vacuum and replaces it with long wavelength, slow, but quite signifi- level, exposing the core electronic structure capabilities of SIESTA to
cant fluctuations. The gap remains sizable, and coherent transport is other programs. It will be necessary to redesign some of the internal
not likely. However, the frontier orbitals evolve in a very suggestive data structures to remove global variables and encapsulate them into
way for enhanced diffusive electron transport.184 objects or derived types associated with particular configurations
and stages of the calculations. This encapsulation will be matched
E. Use of SIESTA in other fields with the streamlining of the input/output operations. This work will
Although an exhaustive summary of all the recent results open the door to the creation of complex workflows leveraging the
obtained with SIESTA is out of the scope of this work, we would strengths of various codes.

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-25

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

Accelerated hybrid architectures (including, for example, FIS2009-12721, FIS2012-37549, FIS2015-64886-P, and RTC-2016-
GPUs) are very likely going to feature prominently in the upcoming 5681-7), the latter one together with Simune Atomistics Ltd. We
exascale machines. In the case of SIESTA, the data indirection associ- are thankful for financial support from the Spanish Ministry of Sci-
ated with the handling of sparse matrices limits the acceleration pos- ence, Innovation and Universities through Grant No. PGC2018-
sibilities of the section of the code that builds the Hamiltonian and 096955-B.
overlap matrices, but the solver stage is more amenable to porting, We acknowledge the Severo Ochoa Center of Excellence Pro-
and in fact, several solver libraries used by SIESTA are being enhanced gram [Grant Nos. SEV-2015-0496 (ICMAB) and SEV-2017-0706
to offer the GPU support, as mentioned in Sec. III G 2. (ICN2)], the GenCat (Grant No. 2017SGR1506), and the European
Modularization and the use of new programming models result Union MaX Center of Excellence (EU-H2020 Grant No. 824143).
in an increase in the complexity of the building and deployment of P.G.-F. acknowledges support from Ramón y Cajal (Grant No.
the code. We will leverage the ESL bundle created to facilitate the RyC-2013-12515). J.I.C. acknowledges Grant No. RTI2018-097895-
use of the modules in the ESL collection, to streamline SIESTA’s build- B-C41.
ing process, and to explore containerization as an option for the R.C. acknowledges the European Union’s Horizon 2020
deployment of the code. Research and Innovation Program under Marie Skłodoswka-Curie
The “pseudopotential barrier to entry” has been lowered by Grant Agreement No. 665919.
the availability of curated databases supporting the PSML format. D.S.P, P.K., and P.B. acknowledge Grant No. MAT2016-78293-
Basis sets are a perennial challenge, but new tools and ideas are C6, FET-Open No. 863098, and UPV-EHU Grant No. IT1246-19.
being explored to provide users with appropriate basis sets: high- V. W. Yu was supported by a MolSSI Fellowship (U.S. NSF
throughput workflows for optimization, “tiers” of quality/cost, but Award No. 1547580), and V.B. and V.W.Y. were supported by
perhaps not just of a simple “Periodic Table” form, as offered by the ELSI Development by the NSF (Award No. 1450280). We also
other codes (e.g., FHI-aims13 ), but with a possible dependence on an acknowledge Honghui Shang and Xinming Qin for giving us access
approximate characterization of the chemical environment in which to the HONPAS code, where a preliminary version of the hybrid
a given atom finds itself. functional support described here was implemented.
Complementary to the underlying basis set optimization that We are indebted to other contributors to the SIESTA project
focuses on providing an adequate variational freedom, an on-the-fly whose names can be seen in the Docs/Contributors.txt file of
contraction of the basis set, which results in a set of lower-cardinality the SIESTA distribution, and we thank those, too many to list, con-
adapted to the description of the occupied subspace, can be exploited tributing fixes, comments, clarifications, and documentation for the
for increased efficiency. This is particularly relevant for FOE meth- code.
ods (see Sec. III G 1) in which the number of polynomial terms
depends on the extent of the spectrum. DATA AVAILABILITY
The original claim to fame of SIESTA was based on its linear- The data that support the findings of this study are available
scaling solver. We are in the process of re-designing the O(N) from the corresponding author upon reasonable request.
code with a new, more efficient backend based on the DBCSR
library for handling distributed block-sparse matrices194,195 with the
REFERENCES
MatrixSwitch library82 acting as an intermediary interface between it
1
and high-level physical ideas and algorithms. A connection between J. M. Soler, E. Artacho, J. D. Gale, A. García, J. Junquera, P. Ordejón, and
the internal SIESTA formats and MatrixSwitch itself has been recently D. Sánchez-Portal, “The SIESTA method for ab initio order-n materials simula-
tion,” J. Phys.: Condens. Matter 14, 2745–2779 (2002).
provided, using initially the cubic-scaling libOMM library196 as a 2
E. Artacho, E. Anglada, O. Diéguez, J. D. Gale, A. García, J. Junquera, R. M.
test bed, hence still using a dense coefficient matrix, as it corre- Martin, P. Ordejón, J. M. Pruneda, D. Sánchez-Portal, and J. M. Soler, “The
sponds to the case without localization constraints in the solu- SIESTA method; developments and applicability,” J. Phys.: Condens. Matter 20,
tion of the electronic structure problem. The implementation of a 064208 (2008).
sparse coefficient matrix will make it possible to perform efficient 3
G. Galli, “Linear scaling methods for electronic structure calculations and quan-
O(N) calculations. The computational effort can be further reduced tum molecular dynamics simulations,” Curr. Opin. Solid State Mater. Sci. 1,
through the analysis of sparsity of the Hamiltonian and overlap 864–874 (1996).
4
matrices and their re-organization in the block-compressed sparse S. Goedecker, “Linear scaling electronic structure methods,” Rev. Mod. Phys. 71,
1085–1123 (1999).
form. 5
P. Ordejón, E. Artacho, and J. M. Soler, “Self-consistent order-n density-
Other developments in the pipeline are linear-response calcula- functional calculations for very large systems,” Phys. Rev. B 53, R10441–R10444
tions for arbitrary distortions, electronic transport calculations with (1996).
spin–orbit coupling, thermal transport with the Green–Kubo for- 6
D. Sánchez-Portal, P. Ordejón, E. Artacho, and J. M. Soler, “Density-functional
malism, as described in Ref. 197, a redesign of the molecular dynam- method for very large systems with LCAO basis sets,” Int. J. Quantum Chem. 65,
ics subsystem, and the development of workflows for the generation 453–461 (1997).
7
of data for SCALE-UP. O. F. Sankey and D. J. Niklewski, “Ab initio multicenter tight-binding model
for molecular-dynamics simulations and other applications in covalent systems,”
Phys. Rev. B 40, 3979–3995 (1989).
8
ACKNOWLEDGMENTS E. Artacho, D. Sánchez-Portal, P. Ordejón, A. García, and J. M. Soler, “Linear-
scaling ab initio calculations for large and complex systems,” Phys. Status Solidi B
SIESTA development was historically supported by different 215, 809–817 (1999).
Spanish National Plan projects (Project Nos. MEC-DGES-PB95- 9
J. Junquera, O. Paz, D. Sánchez-Portal, and E. Artacho, “Numerical atomic
0202, MCyT-BFM2000-1312, MEC-BFM2003-03372, FIS2006-12117, orbitals for linear-scaling calculations,” Phys. Rev. B 64, 235111 (2001).

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-26

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

10 30
E. Anglada, J. M. Soler, J. Junquera, and E. Artacho, “Systematic generation S. L. Dudarev, G. A. Botton, S. Y. Savrasov, C. J. Humphreys, and A. P.
of finite-range atomic basis sets for linear-scaling calculations,” Phys. Rev. B 66, Sutton, “Electron-energy-loss spectra and the structural stability of nickel oxide:
205101 (2002). An LSDA+U study,” Phys. Rev. B 57, 1505–1509 (1998).
11 31
See http://www.openmx-square.org/ for information about the OpenMX code. M. Cococcioni and S. de Gironcoli, “Linear response approach to the calculation
12 of the effective interaction parameters in the LDA+U method,” Phys. Rev. B 71,
See https://www.synopsys.com/silicon/quantumatk.html for information about
the QuantumATK package. 035105 (2005).
13 32
V. Blum, R. Gehrke, F. Hanke, P. Havu, V. Havu, X. Ren, K. Reuter, and R. J. Powell and W. E. Spicer, “Optical properties of NiO and CoO,” Phys. Rev. B
M. Scheffler, “Ab initio molecular simulations with numeric atom-centered 2, 2182–2193 (1970).
33
orbitals,” Comput. Phys. Commun. 180, 2175–2196 (2009). H. A. Alperin, J. Phys. Soc. Jpn. Suppl. B 17, 12 (1962).
14 34
A. Carreras, S. Conejeros, A. Camón, A. García, N. Casañ-Pastor, P. Alemany, A. K. Cheetham and D. A. O. Hope, “Magnetic ordering and exchange effects
and E. Canadell, “Charge delocalization, oxidation states, and silver mobility in in the antiferromagnetic solid solutions Mnx Ni1−x O,” Phys. Rev. B 27, 6964–6967
the mixed silver–copper oxide AgCuO2,” Inorg. Chem. 58, 7026–7035 (2019). (1983).
15 35
M. Brandbyge, J.-L. Mozos, P. Ordejón, J. Taylor, and K. Stokbro, “Density- M. Dion, H. Rydberg, E. Schröder, D. C. Langreth, and B. I. Lundqvist, “van
functional method for nonequilibrium electron transport,” Phys. Rev. B 65, der Waals density functional for general geometries,” Phys. Rev. Lett. 92, 246401
165401 (2002). (2004).
16 36
F. Corsetti, E. Artacho, J. M. Soler, S. S. Alexandre, and M.-V. Fernández- K. Berland and P. Hyldgaard, “Exchange functional that tests the robustness of
Serra, “Room temperature compressibility and diffusivity of liquid water from first the plasmon description of the van der Waals density functional,” Phys. Rev. B 89,
principles,” J. Chem. Phys. 139, 194502 (2013). 035412 (2014).
17 37
X. Gonze, B. Amadon, P.-M. Anglade, J.-M. Beuken, F. Bottin, P. Boulanger, G. Román-Pérez and J. M. Soler, “Efficient implementation of a van der Waals
F. Bruneval, D. Caliste, R. Caracas, M. Côté, T. Deutsch, L. Genovese, P. Ghosez, density functional: Application to double-wall carbon nanotubes,” Phys. Rev. Lett.
M. Giantomassi, S. Goedecker, D. R. Hamann, P. Hermet, F. Jollet, G. Jomard, 103, 096102 (2009).
S. Leroux, M. Mancini, S. Mazevet, M. J. T. Oliveira, G. Onida, Y. Pouillon, 38
O. A. Vydrov and T. Van Voorhis, “Nonlocal van der Waals density functional:
T. Rangel, G.-M. Rignanese, D. Sangalli, R. Shaltaf, M. Torrent, M. J. The simpler the better,” J. Chem. Phys. 133, 244103 (2010).
Verstraete, G. Zerah, and J. W. Zwanziger, “ABINIT: First-principles approach to 39
L. Kong, G. Román-Pérez, J. M. Soler, and D. C. Langreth, “Energetics and
material and nanosystem properties,” Comput. Phys. Commun. 180, 2582–2615 dynamics of h2 adsorbed in a nanoporous material at low temperature,” Phys.
(2009). Rev. Lett. 103, 096103 (2009).
18
See https://www.gnu.org/licenses/gpl-3.0.html for more details about the GPL. 40
H. Gonzalez-Herrero, J. M. Gomez-Rodriguez, P. Mallet, M. Moaied, J. J.
19
See https://launchpad.net/siesta for more details about the Launchpad platform Palacios, C. Salgado, M. M. Ugeda, J.-Y. Veuillen, F. Yndurain, and I. Brihuega,
for Siesta development. “Atomic-scale control of graphene magnetism by using hydrogen atoms,” Science
20
See https://gitlab.com/siesta-project for more details about the Gitlab platform 352, 437–441 (2016).
for Siesta development. 41
J. Wang, G. Román-Pérez, J. M. Soler, E. Artacho, and M.-V. Fernández-
21
A. García, M. J. Verstraete, Y. Pouillon, and J. Junquera, “The PSML format and Serra, “Density, structure, and dynamics of water: The effect of van der Waals
library for norm-conserving pseudopotential data curation and interoperability,” interactions,” J. Chem. Phys. 134, 024516 (2011).
Comput. Phys. Commun. 227, 51–71 (2018). 42
J. Heyd, G. E. Scuseria, and M. Ernzerhof, “Hybrid functionals based on a
22
See https://siesta-project.github.io/psml-docs for more information about the screened Coulomb potential,” J. Chem. Phys. 118, 8207–8215 (2003).
PSML ecosystem; accessed November 2019. 43
J. Heyd, G. E. Scuseria, and M. Ernzerhof, “Erratum: “Hybrid functionals based
23
D. R. Hamann, “Optimized norm-conserving Vanderbilt pseudopotentials,” on a screened Coulomb potential”,” J. Chem. Phys. 124, 219906 (2006).
Phys. Rev. B 88, 085117 (2013). 44
A. V. Krukau, O. A. Vydrov, A. F. Izmaylov, and G. E. Scuseria, “Influence
24
ATOM code for the generation of norm-conserving pseudopotentials. The ver- of the exchange screening parameter on the performance of screened hybrid
sion maintained by the SIESTA project can be accessed at http://icmab.es/siesta/ functionals,” J. Chem. Phys. 125, 224106 (2006).
Pseudopotentials/index.html. An alternative version is available at http://bohr. 45
H. Shang, Z. Li, and J. Yang, “Implementation of screened hybrid density func-
inesc-mn.pt/∼ jlm/pseudo.html; accessed July 2017. tional for periodic systems with numerical atomic orbitals: Basis function fitting
25
M. J. van Setten, M. Giantomassi, E. Bousquet, M. J. Verstraete, D. R. Hamann, and integral screening,” J. Chem. Phys. 135, 034110 (2011).
X. Gonze, and G.-M. Rignanese, “The PSEUDODOJO: Training and grading a 46
See https://sourceforge.net/projects/libint/ for provided by E. Valeev and J. T.
85 element optimized norm-conserving pseudopotential table,” Comput. Phys. Fermann.
Commun. 226, 39–54 (2018). 47
26 S. Obara and A. Saika, “Efficient recursive computation of molecular integrals
See http://www.pseudo-dojo.org for information about the Pseudo-Dojo
over cartesian Gaussian functions,” J. Chem. Phys. 84, 3963–3974 (1986).
project and database. 48
27 M. Head-Gordon and J. A. Pople, “A method for two-electron Gaussian integral
X. Gonze, F. Jollet, F. Abreu Araujo, D. Adams, B. Amadon, T. Applencourt,
and integral derivative evaluation using recurrence relations,” J. Chem. Phys. 89,
C. Audouze, J.-M. Beuken, J. Bieder, A. Bokhanchuk, E. Bousquet, F. Bruneval,
5777–5786 (1988).
D. Caliste, M. Côté, F. Dahm, F. Da Pieve, M. Delaveau, M. Di Gennaro, 49
B. Dorado, C. Espejo, G. Geneste, L. Genovese, A. Gerossier, M. Giantomassi, J. P. Perdew, K. Burke, and M. Ernzerhof, “Generalized gradient approximation
Y. Gillet, D. R. Hamann, L. He, G. Jomard, J. Laflamme Janssen, S. Le Roux, made simple,” Phys. Rev. Lett. 77, 3865–3868 (1996).
50
A. Levitt, A. Lherbier, F. Liu, I. Lukačević, A. Martin, C. Martins, M. J. T. Oliveira, C. Kittel, Introduction to Solid State Physics (John Wiley & Sons, New York,
S. Poncé, Y. Pouillon, T. Rangel, G.-M. Rignanese, A. H. Romero, B. Rousseau, 1986).
51
O. Rubel, A. A. Shukri, M. Stankovski, M. Torrent, M. J. Van Setten, B. Van S. H. Wemple, “Polarization fluctuations and the optical-absorption edge in
Troeye, M. J. Verstraete, D. Waroquiers, J. Wiktor, B. Xu, A. Zhou, and J. W. BaTiO3 ,” Phys. Rev. B 2, 2679–2689 (1970).
Zwanziger, “Recent developments in the ABINIT software package,” Comput. 52
R. Cuadrado and J. I. Cerdá, “Fully relativistic pseudopotential formalism under
Phys. Commun. 205, 106–131 (2016). an atomic orbital basis: Spin–orbit splittings and magnetic anisotropies,” J. Phys.:
28
V. I. Anisimov, J. Zaanen, and O. K. Andersen, “Band theory and Mott Condens. Matter 24, 086005 (2012).
insulators: Hubbard U instead of stoner I,” Phys. Rev. B 44, 943–954 (1991). 53
L. Fernández-Seivane, M. A. Oliveira, S. Sanvito, and J. Ferrer, “On-
29
B. Himmetoglu, A. Floris, S. de Gironcoli, and M. Cococcioni, “Hubbard- site approximation for spin–orbit coupling in linear combination of atomic
corrected DFT energy functionals: The LDA+U description of correlated sys- orbitals density functional methods,” J. Phys.: Condens. Matter 18, 7999–8013
tems,” Int. J. Quantum Chem. 114, 14–49 (2014). (2006).

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-27

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

54 73
F. Fernández-Seivane, M. A. Oliveira, S. Sanvito, and J. Ferrer, “Erratum: W. Dawson and T. Nakajima, “Massively parallel sparse matrix func-
“On-site approximation for spin–orbit coupling in LCAO density functional tion calculations with NTPoly,” Comput. Phys. Commun. 225, 154–165
methods”,” J. Phys.: Condens. Matter 19, 489001 (2007). (2018).
55 74
L. A. Hemstreet, C. Y. Fong, and J. S. Nelson, “First-principles calculations of A. Tsolakidis, D. Sánchez-Portal, and R. M. Martin, “Calculation of the optical
spin-orbit splittings in solids using nonlocal separable pseudopotentials,” Phys. response of atomic clusters using time-dependent density functional theory and
Rev. B 47, 4238 (1993). local orbitals,” Phys. Rev. B 66, 235416 (2002).
56 75
F. Zirkelbach, P.-Y. Prodhomme, P. Han, R. Cherian, and G. Bester, “Large- E. Artacho and D. D. O’Regan, “Quantum mechanics in an evolving Hilbert
scale atomic effective pseudopotential program including an efficient spin-orbit space,” Phys. Rev. B 95, 115155 (2017).
coupling treatment in real space,” Phys. Rev. B 91, 075119 (2015). 76
J. K. Tomfohr and O. F. Sankey, “Time-dependent simulation of conduction
57
V. W.-z. Yu, F. Corsetti, A. García, W. P. Huhn, M. Jacquelin, W. Jia, B. Lange, through a molecule,” Phys. Status Solidi B 226, 115–123 (2001).
L. Lin, J. Lu, W. Mi, A. Seifitokaldani, Á. Vázquez-Mayagoitia, C. Yang, H. Yang, 77
A. A. Correa, J. Kohanoff, E. Artacho, D. Sánchez-Portal, and A. Caro, “Nona-
and V. Blum, “ELSI: A unified software interface for Kohn-Sham electronic diabatic forces in ion-solid interactions: The initial stages of radiation damage,”
structure solvers,” Comput. Phys. Commun. 222, 267–285 (2018). Phys. Rev. Lett. 108, 213201 (2012).
58
V. W. zhe Yu, C. Campos, W. Dawson, A. García, V. Havu, B. Hourahine, W. P. 78
M. A. Zeb, J. Kohanoff, D. Sánchez-Portal, A. Arnau, J. I. Juaristi, and E.
Huhn, M. Jacquelin, W. Jia, M. Keçeli, R. Laasner, Y. Li, L. Lin, J. Lu, J. Moussa, Artacho, “Electronic stopping power in gold: The role of d electrons and the H/He
J. E. Roman, Á. Vázquez-Mayagoitia, C. Yang, and V. Blum, “ELSI—An open anomaly,” Phys. Rev. Lett. 108, 225504 (2012).
infrastructure for electronic structure solvers,” arXiv:1912.13403 [physics.comp- 79
R. Ullah, F. Corsetti, D. Sánchez-Portal, and E. Artacho, “Electronic stopping
ph] (2019). power in a narrow band gap semiconductor from first principles,” Phys. Rev. B
59
J. Choi, J. Demmel, I. Dhillon, J. Dongarra, S. Ostrouchov, A. Petitet, K. Stanley, 91, 125203 (2015).
D. Walker, and R. C. Whaley, “SCALAPACK: A portable linear algebra library for 80
J. Halliday and E. Artacho, “Anisotropy of electronic stopping power in
distributed memory computers—Design issues and performance,” Comput. Phys. graphite,” Phys. Rev. B 100, 104112 (2019).
Commun. 97, 1–15 (1996). 81
60 F. Corsetti, “Performance analysis of electronic structure codes on HPC systems:
T. Auckenthaler, V. Blum, H.-J. Bungartz, T. Huckle, R. Johanni, L. Krämer, A case study of siesta,” PLoS One 9, 1–8 (2014).
B. Lang, H. Lederer, and P. R. Willems, “Parallel solution of partial symmetric 82
F. Corsetti, https://gitlab.e-cam2020.eu/esl/omm.
eigenvalue problems from electronic structure calculations,” Parallel Comput. 37, 83
783–794 (2011). E. Anderson, A. Benzoni, J. Dongarra, S. Moulton, S. Ostrouchov, B.
61 Tourancheau, and R. van de Geijn, “Basic linear algebra communication sub-
A. Marek, V. Blum, R. Johanni, V. Havu, B. Lang, T. Auckenthaler, A. Heinecke,
programs,” in 1991 Proceedings of the Sixth Distributed Memory Computing
H.-J. Bungartz, and H. Lederer, “The ELPA library: Scalable parallel eigenvalue
Conference (IEEE, 1991), pp. 287–290.
solutions for electronic structure theory and computational science,” J. Phys.: 84
Condens. Matter 26, 213201 (2014). L. S. Blackford, J. Choi, A. Cleary, E. D’Azevedo, J. Demmel, I. Dhillon, J.
62 Dongarra, S. Hammarling, G. Henry, A. Petitet, K. Stanley, D. Walker, and
P. Kuṡ, A. Marek, S. Koecher, H.-H. Kowalski, C. Carbogno, C. Scheurer,
R. C. Whaley, ScaLAPACK Users’ Guide (Society for Industrial and Applied
K. Reuter, M. Scheffler, and H. Lederer, “Optimizations of the eigensolvers in the
Mathematics, Philadelphia, PA, 1997).
ELPA library,” Parallel Comput. 85, 167–177 (2019). 85
63 F. Tisseur and J. Dongarra, “A parallel divide and conquer algorithm for the
S. Goedecker, “Integral representation of the fermi distribution and its appli-
symmetric eigenvalue problem on distributed memory architectures,” SIAM J. Sci.
cations in electronic-structure calculations,” Phys. Rev. B 48, 17573–17575
Comput. 20, 2223–2236 (1999).
(1993). 86
64 T. N. Todorov, “Time-dependent tight binding,” J. Phys.: Condens. Matter 13,
S. Mohr, W. Dawson, M. Wagner, D. Caliste, T. Nakajima, and L. Genovese,
10125–10148 (2001).
“Efficient computation of sparse matrix functions for large-scale electronic struc- 87
ture calculations: The CheSS library,” J. Chem. Theory Comput. 13, 4684–4698 I. Tavernelli, U. F. Röhrig, and U. Rothlisberger, “Molecular dynamics in elec-
(2017). tronically excited states using time-dependent density functional theory,” Mol.
65
L. Genovese, A. Neelov, S. Goedecker, T. Deutsch, S. A. Ghasemi, A. Willand, Phys. 103, 963–981 (2005).
88
D. Caliste, O. Zilberberg, M. Rayson, A. Bergman, and R. Schneider, “Daubechies CPMD v3.13, Copyright IBM Corp (1990–2008), copyright MPI fuer Festkoer-
wavelets as a basis set for density functional pseudopotential calculations,” perforschung Stuttgart (1997–2001).
89
J. Chem. Phys. 129, 014109 (2008). P. López-Tarifa, M.-A. Hervé du Penhoat, R. Vuilleumier, M.-P. Gaigeot,
66 I. Tavernelli, A. Le Padellec, J.-P. Champeaux, M. Alcamí, P. Moretto-Capelle,
L. Lin, M. Chen, C. Yang, and L. He, “Accelerating atomic orbital-based elec-
tronic structure calculation via pole expansion and selected inversion,” J. Phys.: F. Martín, and M.-F. Politis, “Ultrafast nonadiabatic fragmentation dynam-
Condens. Matter 25, 295501 (2013). ics of doubly charged uracil in a gas phase,” Phys. Rev. Lett. 107, 023202
67
L. Lin, A. García, G. Huhs, and C. Yang, “SIESTA-PEXSI: Massively parallel (2011).
90
method for efficient and accurate ab initio materials simulation without matrix P. López-Tarifa, M.-A. Penhoat, R. Vuilleumier, M.-P. Gaigeot, U. Rothlis-
diagonalization,” J. Phys.: Condens. Matter 26, 305503 (2014). berger, I. Tavernelli, A. Le Padellec, J.-P. Champeaux, M. Alcami, P. Moretto
68
W. Hu, L. Lin, C. Yang, and J. Yang, “Electronic structure and aromatic- Capelle, F. Martín, and M. Politis, “Time-dependent density functional theory
ity of large-scale hexagonal graphene nanoflakes,” J. Chem. Phys. 141, 214704 molecular dynamics simulation of doubly charged uracil in gas phase,” Cent. Eur.
(2014). J. Phys. 12, 97–102 (2014).
69 91
F. Corsetti, “The orbital minimization method for electronic structure calcula- M.-P. Gaigeot, P. Lopez-Tarifa, F. Martin, M. Alcami, R. Vuilleumier, I.
tions with finite-range atomic basis sets,” Comput. Phys. Commun. 185, 873–883 Tavernelli, M.-A. Hervé du Penhoat, and M.-F. Politis, “Theoretical investigation
(2014). of the ultrafast dissociation of ionised biomolecules immersed in water: Direct and
70
T. Imamura, S. Yamada, and M. Machida, “Development of a high-performance indirect effects,” Mutat. Res. 704, 45–53 (2010).
92
eigensolver on a peta-scale next-generation supercomputer system,” Prog. Nucl. A. D. Becke, “Density-functional exchange-energy approximation with correct
Sci. Technol. 2, 643–650 (2011). asymptotic behavior,” Phys. Rev. A 38, 3098–3100 (1988).
71 93
J. Dongarra, M. Gates, A. Haidar, J. Kurzak, P. Luszczek, S. Tomov, and C. Lee, W. Yang, and R. G. Parr, “Development of the Colle-Salvetti correlation-
I. Yamazaki, “Accelerating numerical dense linear algebra calculations with energy formula into a functional of the electron density,” Phys. Rev. B 37, 785–789
GPUs,” in Numerical Computations with GPUs (Springer, Cham, 2014), pp. 1–26. (1988).
72 94
V. Hernandez, J. E. Roman, and V. Vidal, “SLEPc: A scalable and flexible toolkit J. M. Pruneda, D. Sánchez-Portal, A. Arnau, J. I. Juaristi, and E. Artacho, “Elec-
for the solution of eigenvalue problems,” ACM Trans. Math. Software 31, 351–362 tronic stopping power in LiF from first principles,” Phys. Rev. Lett. 99, 235501
(2005). (2007).

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-28

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

95 118
A. A. Correa, “Calculating electronic stopping power in materials from first S. Das, Y. L. Tang, Z. Hong, M. A. P. Gonçalves, M. R. McCarter, C. Klewe,
principles,” Comput. Mater. Sci. 150, 291–303 (2018). K. X. Nguyen, F. Gómez-Ortiz, P. Shafer, E. Arenholz, V. A. Stoica, S.-L. Hsu,
96 B. Wang, C. Ophus, J. F. Liu, C. T. Nelson, S. Saremi, B. Prasad, A. B. Mei,
J. M. Pruneda, S. K. Estreicher, J. Junquera, J. Ferrer, and P. Ordejón, “Ab initio
local vibrational modes of light impurities in silicon,” Phys. Rev. B 65, 075210 D. G. Schlom, J. Íñiguez, P. García-Fernández, D. A. Muller, L. Q. Chen, J.
(2002). Junquera, L. W. Martin, and R. Ramesh, “Observation of room-temperature polar
97
N. Papior, N. Lorente, T. Frederiksen, A. García, and M. Brandbyge, “Improve- skyrmions,” Nature 568, 368–372 (2019).
119
ments on non-equilibrium and transport green function techniques: The next- A. K. Yadav, K. X. Nguyen, Z. Hong, P. García-Fernández, P. Aguado-Puente,
generation transiesta,” Comput. Phys. Commun. 212, 8–24 (2017). C. T. Nelson, S. Das, B. Prasad, D. Kwon, S. Cheema, A. I. Khan, C. Hu, J. Íñiguez,
98
N. Papior, G. Calogero, S. Leitherer, and M. Brandbyge, “Removing all periodic J. Junquera, L.-Q. Chen, D. A. Muller, R. Ramesh, and S. Salahuddin, “Spatially
boundary conditions: Efficient nonequilibrium Green’s function calculations,” resolved steady-state negative capacitance,” Nature 565, 468–471 (2019).
120
Phys. Rev. B 100, 195417 (2019). R. Ierusalimschy, Programming in Lua, 4th ed. (Feisty Duck Digital Book
99
K. W. Jacobsen, J. T. Falkenberg, N. Papior, P. Bøggild, A.-P. Jauho, and Distribution, 2016).
121
M. Brandbyge, “All-graphene edge contacts: Electrical resistance of graphene See https://siesta-project.github.io/flos/ldoc/index.html for more information
t-junctions,” Carbon 101, 101–106 (2016). about the library of Lua scripts for use with Siesta.
100 122
M. Kolmer, P. Brandimarte, J. Lis, R. Zuzak, S. Godlewski, H. Kawai, A. S. Smidstrup, A. Pedersen, K. Stokbro, and H. Jónsson, “Improved ini-
Garcia-Lekue, N. Lorente, T. Frederiksen, C. Joachim, D. Sanchez-Portal, and tial guess for minimum energy path calculations,” J. Chem. Phys. 140, 214106
M. Szymonski, “Electronic transport in planar atomic-scale structures measured (2014).
123
by two-probe scanning tunneling spectroscopy,” Nat. Commun. 10, 1573 (2019). D. Sheppard, R. Terrell, and G. Henkelman, “Optimization methods for
101
P. Brandimarte, M. Engelund, N. Papior, A. Garcia-Lekue, T. Frederiksen, finding minimum energy paths,” J. Chem. Phys. 128, 134106 (2008).
124
and D. Sánchez-Portal, “A tunable electronic beam splitter realized with crossed S. A. Trygubenko and D. J. Wales, “A doubly nudged elastic band method for
graphene nanoribbons,” J. Chem. Phys. 146, 092318 (2017). finding transition states,” J. Chem. Phys. 120, 2082–2094 (2004).
102 125
N. Papior, SISL: v0.9.7, 2019. G.-R. Qian, X. Dong, X.-F. Zhou, Y. Tian, A. R. Oganov, and H.-T. Wang,
103
See https://github.com/tfrederiksen/inelastica/ for more information about the “Variable cell nudged elastic band method for studying solid-solid structural phase
Inelastica package. transitions,” Comput. Phys. Commun. 184, 2111–2118 (2013).
104 126
T. Frederiksen, M. Paulsson, M. Brandbyge, and A.-P. Jauho, “Inelastic trans- See https://esl.cecam.org/Flook for more details about the Fortran-Lua bridge,
port theory from first principles: Methodology and application to nanoscale including interoperability of data structures.
devices,” Phys. Rev. B 75, 205413 (2007). 127
G. Pizzi, A. Cepellotti, R. Sabatini, N. Marzari, and B. Kozinsky, “Aiida:
105
See http://www.wannier.org for information about the Wannier90 code. Automated interactive infrastructure and database for computational science,”
106
G. Pizzi, V. Vitale, R. Arita, S. Blügel, F. Freimuth, G. Géranton, M. Comput. Mater. Sci. 111, 218–230 (2016).
128
Gibertini, D. Gresch, C. Johnson, T. Koretsune, J. Ibañez-Azpiroz, H. Lee, J.-M. See http://www.aiida.net/ for more information about the AiiDA platform.
Lihm, D. Marchand, A. Marrazzo, Y. Mokrousov, J. I. Mustafa, Y. Nohara, 129
See https://aiida.readthedocs.io/ for full details of the implementation and
Y. Nomura, L. Paulatto, S. Poncé, T. Ponweiser, J. Qiao, F. Thöle, S. S. Tsirkin, usage of AiiDA.
M. Wierzbowska, N. Marzari, D. Vanderbilt, I. Souza, A. A. Mostofi, and J. R. 130
See https://aiida-siesta-plugin.readthedocs.io/ for documentation on the use of
Yates, “Wannier90 as a community code: New features and applications,” J. Phys.: the AiiDA-Siesta plugin.
Condens. Matter 32, 165902 (2020). 131
107
S. G. Mayo, F. Yndurain, and J. M. Soler, “Band unfolding made simple,”
N. Marzari and D. Vanderbilt, “Maximally localized generalized wannier J. Phys.: Condens. Matter 32, 205902 (2020).
functions for composite energy bands,” Phys. Rev. B 56, 12847–12865 (1997). 132
108
G. Calogero, N. Papior, M. Koleini, M. H. L. Larsen, and M. Brandbyge, “Multi-
N. Marzari, A. A. Mostofi, J. R. Yates, I. Souza, and D. Vanderbilt, “Maximally scale approach to first-principles electron transport beyond 100 nm,” Nanoscale
localized wannier functions: Theory and applications,” Rev. Mod. Phys. 84, 1419– 11, 6153–6164 (2019).
1475 (2012). 133
109
A. H. Larsen, J. J. Mortensen, J. Blomqvist, I. E. Castelli, R. Christensen,
D. Vanderbilt, Berry Phases in Electronic Structure Theory (Cambridge Univer- M. Dułak, J. Friis, M. N. Groves, B. Hammer, C. Hargus, E. D. Hermes, P. C.
sity Press, 2018). Jennings, P. B. Jensen, J. Kermode, J. R. Kitchin, E. L. Kolsbjerg, J. Kubal,
110
See https://github.com/romerogroup/DMFTwDFT for information about the K. Kaasbjerg, S. Lysgaard, J. B. Maronsson, T. Maxson, T. Olsen, L. Pastewka, A.
DMFTwDFT code. Peterson, C. Rostgaard, J. Schiøtz, O. Schütt, M. Strange, K. S. Thygesen, T. Vegge,
111
V. Singh, U. Herath, B. Wah, X. Liao, A. H. Romero, and H. Park, L. Vilhelmsen, M. Walter, Z. Zeng, and K. W. Jacobsen, “The atomic simula-
“DMFTWDFT: An open-source code combining dynamical mean field theory tion environment—A python library for working with atoms,” J. Phys.: Condens.
with various density functional theory packages,” arXiv:2002.00068v1 (2020). Matter 29, 273002 (2017).
112 134
X. Wu, A. Selloni, and R. Car, “Order-n implementation of exact exchange in A. Kokalj, “XCrySDen: A crystalline and molecular structure visualisation
extended insulating systems,” Phys. Rev. B 79, 085102 (2009). program,” http://www.xcrysden.org.
113 135
P. García-Fernández, J. C. Wojdeł, J. Íñiguez, and J. Junquera, “Second- K. Momma, “VESTA: A 3D visualization program for structural models,
principles method for materials simulations including electron and lattice degrees volumetric data, and crystal morphologies,” http://jp-minerals.org/vesta/en/.
of freedom,” Phys. Rev. B 93, 195137 (2016). 136
A. V. Postnikov and N. B. Mortazavi Amiri, “Calculated vibration spectrum
114
J. C. Wojdeł, P. Hermet, M. P. Ljungberg, P. Ghosez, and J. Íñiguez, “First- of monoclinic Cu2 SnSe3 in comparison with kesterite-type Cu2 ZnSnSe4 ,” Phys.
principles model potentials for lattice-dynamical studies: General methodology Status Solidi A 210, 1332–1335 (2013).
and example of application to ferroic perovskite oxides,” J. Phys.: Condens. Matter 137
M. N. Rao, D. Lamago, A. Ivanov, M. d’Astuto, A. V. Postnikov, R. Hajj Hus-
25, 305401 (2013). sein, T. Basak, S. L. Chaplot, F. Firszt, W. Paszkowicz, S. K. Deb, and O. Pagés,
115
P. Torres, J. A. Seijas-Bellido, C. Escorihuela-Sayalero, J. Íñiguez, and R. Rurali, “Lattice dynamics of the model percolation-type (Zn,Be)Se alloy: Inelastic neutron
“Theoretical investigation of lattice thermal conductivity and electrophononic scattering, ab initio study, and shell-model calculations,” Phys. Rev. B 89, 155201
effects in SrTiO3 ,” Phys. Rev. Mater. 3, 044404 (2019). (2014).
116 138
J. A. Seijas-Bellido, J. Íñiguez, and R. Rurali, “Anisotropy-driven thermal con- Bilbao Crystallographic Server → Raman and Hyper-Raman scattering
ductivity switching and thermal hysteresis in a ferroelectric,” Appl. Phys. Lett. 115, → Spectral Active Modes, https://www.cryst.ehu.es/rep/sam.html.
192903 (2019). 139
V. Kosyak, N. B. Mortazavi Amiri, A. V. Postnikov, and M. A. Scarpulla,
117
J. A. Seijas-Bellido, H. Aramberri, J. Íñiguez, and R. Rurali, “Electric control of “Model of native point defect equilibrium in Cu2 ZnSnS4 and application to
the heat flux through electrophononic effects,” Phys. Rev. B 97, 184306 (2018). one-zone annealing,” J. Appl. Phys. 114, 124501 (2013).

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-29

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

140
N. B. Mortazavi Amiri, “Relation entre motifs structuraux et dynamique de A. M. Minor, L. Q. Chen, J. F. Scott, L. W. Martin, and R. Ramesh, “Observation
réseau dans les cristaux mixtes Cu-Zn-Sn-Se: Étude premiers principes, Ph.D. of polar vortices in oxide superlattices,” Nature 530, 198–201 (2016).
thesis, Université de Lorraine, 2013. 163
P. Shafer, P. García-Fernández, P. Aguado-Puente, A. R. Damodaran, A. K.
141
M. P. Allen and D. J. Tildesley, Computer Simulation of Liquids (Oxford Yadav, C. T. Nelson, S.-L. Hsu, J. C. Wojdeł, J. Íñiguez, L. W. Martin, E. Arenholz,
University Press, Oxford, 1987). J. Junquera, and R. Ramesh, “Emergent chirality in the electric polarization texture
142
A. V. Postnikov, “Vibrations in solids and small particles from first-principles of titanate superlattices,” Proc. Natl. Acad. Sci. U. S. A. 115, 915–920 (2018).
164
calculations,” in Computational Materials Science, NATO Science Series III (Com- D. W. Boukhvalov, M. I. Katsnelson, and A. I. Lichtenstein, “Hydrogen on
puter and System Science) Vol. 187, edited by C. R. A. Catlow and E. Kotomin graphene: Electronic structure, total energy, structural distortions and magnetism
(IOS Press, 2003), pp. 153–166 [Proceedings of the NATO Advanced Study from first-principles calculations,” Phys. Rev. B 77, 035427 (2008).
Institute “Computational Materials Science,” Il Ciocco, September 9–22, 2001]. 165
O. V. Yazyev and L. Helm, “Defect-induced magnetism in graphene,” Phys.
143
O. Coulaud, P. Bordat, P. Fayon, V. Le Bris, I. Baraille, and R. Brown, “Exten- Rev. B 75, 125408 (2007).
sions of the siesta DFT code for simulation of molecules,” Research Report No. 166
M. Slota, A. Keerthi, W. K. Myers, E. Tretyakov, M. Baumgarten, A. Ardavan,
RR-8221, INRIA, 2013. H. Sadeghi, C. J. Lambert, A. Narita, K. Müllen, and L. Bogani, “Magnetic edge
144
P. Koval, M. Barbry, and D. Sánchez-Portal, “PySCF-Nao: An efficient and flex- states and coherent manipulation of graphene nanoribbons,” Nature 557, 691–695
ible implementation of linear response time-dependent density functional theory (2018).
with numerical atomic orbitals,” Comput. Phys. Commun. 236, 188–204 (2019). 167
O. V. Yazyev and S. G. Louie, “Electronic transport in polycrystalline
145
M. Petersilka, U. J. Gossmann, and E. K. U. Gross, “Excitation energies from graphene,” Nat. Mater. 9, 806–809 (2010).
time-dependent density-functional theory,” Phys. Rev. Lett. 76, 1212–1215 (1996). 168
O. V. Yazyev and S. G. Louie, “Topological defects in graphene: Dislocations
146
P. Koval, F. Marchesin, D. Foerster, and D. Sánchez-Portal, “Optical response and grain boundaries,” Phys. Rev. B 81, 195420 (2010).
of silver clusters and their hollow shells from linear-response TDDFT,” J. Phys.: 169
W. Y. Kim and K. S. Kim, “Prediction of very large values of magnetoresistance
Condens. Matter 28, 214001 (2016). in a graphene nanoribbon device,” Nat. Nanotechnol. 3, 408–412 (2008).
147 170
Y. Li and C. A. Ullrich, “The particle–hole map: A computational tool to C. Moreno, M. Vilas-Varela, B. Kretz, A. Garcia-Lekue, M. V. Costache,
visualize electronic excitations,” J. Chem. Theory Comput. 11, 5838–5852 (2015). M. Paradinas, M. Panighel, G. Ceballos, S. O. Valenzuela, D. Peña, and
148
R. Dronskowski and P. E. Bloechl, “Crystal orbital Hamilton populations A. Mugarza, “Bottom-up synthesis of multifunctional nanoporous graphene,”
(COHP): Energy-resolved visualization of chemical bonding in solids based on Science 360, 199–203 (2018).
density-functional calculations,” J. Phys. Chem. 97, 8617–8624 (1993). 171
W. Hu, Y. Huang, X. Qin, L. Lin, E. Kan, X. Li, C. Yang, and J. Yang,
149
M. Barbry, P. Koval, F. Marchesin, R. Esteban, A. G. Borisov, J. Aizpurua, and “Room-temperature magnetism and tunable energy gaps in edge-passivated
D. Sánchez-Portal, “Atomistic near-field nanoplasmonics: Reaching atomic-scale zigzag graphene quantum dots,” npj 2D Mater. Appl. 3, 17 (2019).
resolution in nanooptics,” Nano Lett. 15, 3410–3419 (2015). 172
S. Kim, J. Ihm, H. J. Choi, and Y.-W. Son, “Origin of anomalous electronic
150
F. Marchesin, P. Koval, M. Barbry, J. Aizpurua, and D. Sánchez-Portal, “Optical structures of epitaxial graphene on silicon carbide,” Phys. Rev. Lett. 100, 176802
response of metallic nanojunctions driven by single atom motion,” ACS Photonics (2008).
3, 269–277 (2016). 173
H. S. S. Ramakrishna Matte, A. Gomathi, A. K. Manna, D. J. Late, R. Datta, S. K.
151
M. Barbry, “Plasmons in nanoparticles: Atomistic ab initio theory for large sys- Pati, and C. N. R. Rao, “MoS2 and WS2 analogues of graphene,” Angew. Chem.,
tems, Ph.D. thesis, University of Basque Country, Donostia-San Sebastián, Spain, Int. Ed. 49, 4059–4062 (2010).
2018, http://cfm.ehu.es/view/files/MArc_barbry_2-1.pdf; accessed 14 June 2019. 174
I. Popov, G. Seifert, and D. Tománek, “Designing electrical contacts to mos2
152
E. Lampin, P. L. Palla, P.-A. Francioso, and F. Cleri, “Thermal conductivity monolayers: A computational study,” Phys. Rev. Lett. 108, 156802 (2012).
from approach-to-equilibrium molecular dynamics,” J. Appl. Phys. 114, 033525 175
H. Liu, A. T. Neal, Z. Zhu, Z. Luo, X. Xu, D. Tománek, and P. D. Ye, “Phospho-
(2013). rene: An unexplored 2D semiconductor with a high hole mobility,” ACS Nano 8,
153
S. Illera, M. Pruneda, L. Colombo, and P. Ordejón, “Thermal and trans- 4033–4041 (2014).
port properties of pristine single-layer hexagonal boron nitride: A first principles 176
J. Guan, Z. Zhu, and D. Tománek, “Phase coexistence and metal-insulator tran-
investigation,” Phys. Rev. Mater. 1, 044006 (2017). sition in few-layer phosphorene: A computational study,” Phys. Rev. Lett. 113,
154
S. García-Gil, A. García, and P. Ordejón, “Calculation of core level shifts within 046804 (2014).
DFT using pseudopotentials and localized basis sets,” Eur. Phys. J. B 85, 239 177
W. Hu, L. Lin, C. Yang, J. Dai, and J. Yang, “Edge-modified phosphorene
(2012). nanoflake heterojunctions as highly efficient solar cells,” Nano Lett. 16, 1675–1682
155
S. Garcia-Gil, A. Arnau, and A. Garcia-Lekue, “Exploring large O 1s and N (2016).
1s core level shifts due to intermolecular hydrogen bond formation in organic 178
I. B. Guster, “A bird’s-eye view of charge and spin density wave from first
molecules,” Surf. Sci. 613, 102–107 (2013). principles calculations,” Ph.D. thesis, Universitat Autonoma de Barcelona, 2019.
156 179
See https://en.wikipedia.org/wiki/Reference_counting for more information B. Guster, C. Rubio-Verdú, R. Robles, J. Zaldívar, P. Dreher, M. Pruneda, J. Á.
about the reference-counting approach to memory management. Silva-Guillén, D.-J. Choi, J. I. Pascual, M. M. Ugeda, P. Ordejón, and E. Canadell,
157
See https://www.max-centre.eu/ for more information about the MaX (Mate- “Coexistence of elastic modulations in the charge density wave state of 2H-NbSe2 ,”
rials at the eXascale) EU Center of Excellence. Nano Lett. 19, 3027–3032 (2019).
158 180
See https://esl.cecam.org/ for information about the Electronic Structure B. Guster, M. Pruneda, P. Ordejón, E. Canadell, and J.-P. Pouget, “Evidence for
Library initiative. the weak coupling scenario of the Peierls transition in the blue bronze,” Phys. Rev.
159
See https://l_sim.gitlab.io/psolver/ for information about the Psolver library Mater. 3, 055001 (2019).
for the solution of the Poisson equation. 181
P. J. de Pablo, F. Moreno-Herrero, J. Colchero, J. Gómez Herrero, P. Herrero,
160
P. Zubko, N. Jecklin, A. Torres-Pardo, P. Aguado-Puente, A. Gloter, C. Licht- A. M. Baró, P. Ordejón, J. M. Soler, and E. Artacho, “Absence of dc-conductivity
ensteiger, J. Junquera, O. Stéphan, and J.-M. Triscone, “Electrostatic coupling and in λ-DNA,” Phys. Rev. Lett. 85, 4992–4995 (2000).
local structural distortions at interfaces in ferroelectric/paraelectric superlattices,” 182
A. Crespo, D. A. Scherlis, M. A. Martí, P. Ordejón, A. E. Roitberg, and D. A.
Nano Lett. 12, 2846–2851 (2012). Estrin, “A DFT-based QM-MM approach designed for the treatment of large
161
P. Aguado-Puente and J. Junquera, “Structural and energetic properties of molecular systems: Application to chorismate mutase,” J. Phys. Chem. B 107,
domains in pbtio3 /srtio3 superlattices from first principles,” Phys. Rev. B 85, 13728–13736 (2003).
184105 (2012). 183
C. F. Sanz-Navarro, R. Grima, A. García, E. A. Bea, A. Soba, J. M. Cela, and
162
A. K. Yadav, C. T. Nelson, S. L. Hsu, Z. Hong, J. D. Clarkson, C. M. Schlepütz, P. Ordejón, “An efficient implementation of a QM-MM method in siesta,” Theor.
A. R. Damodaran, P. Shafer, E. Arenholz, L. R. Dedon, D. Chen, A. Vishwanath, Chem. Acc. 128, 825–833 (2011).

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-30

Published under license by AIP Publishing
 The Journal
ARTICLE scitation.org/journal/jcp
of Chemical Physics

184
G. T. Feliciano, A. J. R. da Silva, G. Reguera, and E. Artacho, “Molecular and edited by C. J. Barrios Hernández, I. Gitler, and J. Klapp (Springer International
electronic structure of the peptide subunit of Geobacter sulfurreducens conductive Publishing, Cham, 2017), pp. 329–339.
pili from first principles,” J. Phys. Chem. A 116, 8023–8030 (2012). 191
R. Escribano and G. M. Muñoz Caro, “Introduction to spectroscopy and
185
D. J. Cole and N. D. M. Hine, “Applications of large-scale density functional astronomical observations,” in Laboratory Astrophysics, edited by G. M. Muñoz
theory in biology,” J. Phys.: Condens. Matter 28, 393001 (2016). Caro and R. Escribano (Springer International Publishing, Cham, 2018),
186 pp. 27–47.
M. Darvish Ganji, “Amino acids interacting with defected carbon nanotubes:
192
Ab initio calculations,” J. Pharm. Health Sci. 4, 157–166 (2016). F. Viñes, J. R. B. Gomes, and F. Illas, “Understanding the reactivity of metallic
187 nanoparticles: Beyond the extended surface model for catalysis,” Chem. Soc. Rev.
W. Li, K. Kotsis, and S. Manzhos, “Comparative density functional theory and
density functional tight binding study of arginine and arginine-rich cell pene- 43, 4922–4939 (2014).
193
trating peptide TAT adsorption on anatase TiO2 ,” Phys. Chem. Chem. Phys. 18, F. Tao, W. Schneider, and P. Kamat, Heterogeneous Catalysis at Nanoscale for
19902–19917 (2016). Energy Applications (Wiley, 2015).
188 194
A. Hermann, “Chemical bonding at high pressure,” in Reviews in Com- See https://www.cp2k.org/dbcsr for information about the DBCSR library for
putational Chemistry (John Wiley & Sons, Ltd., 2017), Chap. 1, pp. 1–41, optimized sparse-matrix multiplication.
https://onlinelibrary.wiley.com/doi/pdf/10.1002/9781119356059.ch1. 195
I. Sivkov, P. Seewald, A. Lazzaro, and J. Hutter, “DBCSR: A blocked sparse
189
T. Liu, E. Artacho, F. Gázquez, G. Walters, and D. Hodell, “Prediction of tensor algebra library,” Parallel Computing: Technology Trends, edited by G.
equilibrium isotopic fractionation of the gypsum/bassanite/water system using R. Joubert, Vol. 36, pp. 331–340.
196
first-principles calculations,” Geochim. Cosmochim. Ac. 244, 1–11 (2019). See http://esl.cecam.org/libOMM for information about the libOMM library.
190 197
E. Mayoral, A. Rey, J. Klapp, A. Gómez, and M. Mayoral, “Ab initio DFT A. Marcolongo, P. Umari, and S. Baroni, “Microscopic theory and quantum
calculations for materials in nuclear research,” in High Performance Computing, simulation of atomic heat transport,” Nat. Phys. 12, 80–84 (2016).

J. Chem. Phys. 152, 204108 (2020); doi: 10.1063/5.0005077 152, 204108-31

Published under license by AIP Publishing